I like @deedydas's work but but this take misses context Sarvam-M isn’t a vanity fine-tune; it’s India’s first open-weights 24 B Indic-centric LLM built under brutal GPU & data scarcity. Judging it by few hours of HuggingFace stats badly misses the point. Most people outside India don't appreciate that compute is quite the invisible ceiling - H100 clusters are still not commercially stocked in India - US export caps tightening next week will squeeze supply even further - Indian teams literally queue for hours of A100/H100 time that US & CN labs get on tap Data is also the long-tail problem Indic languages form <0.01 % of CommonCrawl. You read that right—two orders of magnitude less than Chinese or Spanish. Any local lab must build its corpus first, then train. That’s months of ETL before the first gradient step. Synthetic data is GPU-constrained. Talent pipeline is still forming HPC + RLHF + compiler-level optimisation is new ground in India; Sarvam’s run has already up-skilled dozens of engineers who now know how to wrangle 10 k GPU-hours, FP8 PTQ & GRPO reward engines. Their detailed blog post democratizes a lot of this learning. You can’t AWS-credit your way to that muscle memory. What Sarvam actually shipped - 3.7 M high-diversity Indic prompts, deduped & quality-scored - Two-phase non-think/think alignment that adds+2 pp on IndicGen - GRPO RL with partial-credit rewards—LiveCodeBench jump 0.23→0.44 - FP8 + look-ahead decoding: 2× tokens/s, ½ $/M tok on H100 That means a 🇮🇳-hosted midsize model now matches Gemma-3 27 B and Llama-3.3 70 B on Indic reasoning while costing a fraction to serve. That’s some engineering leverage & definitely not hype. Model adoption is anyways a long-tail - one needs to ship multiple models of non-frontier quality to eventually be able to get to the one that's truly at the frontier (at least along dimensions that we care about). Plus, there's a whole host of Indic-language use-cases where this sovereign model would work much better compared to using any other open-weights model. Look at (LiveCodeBench 0.23→0.44, 2× tokens/s) If you ask for stats, you'll learn that some of their conversational AI platform reaches out to about 50M+ people in just a week. What's next possibly? - Maybe we all recognize the data problem & do Nation-scale data-collection drives (something like CommonCrawl-IN) - Public RL-as-a-service clusters so smaller labs can replicate GRPO - For devs wanting to push the Indic NLP forward, consider forking Sarvam-M, fine-tuning on your domain corpus, benchmarking on Indic-Eval, contributing back patches. Each derivative model widens the knowledge base & closes the English–Indic gap. In summary, celebrating Sarvam's work (I'm not an investor) isn't nationalism, it's recognizing an innovation feat under constraints - India can't out-GPU Mountain View today but there's technical merit on display here, regardless of the metrics. 👏 @pratykumar, @AashaySachdeva, @HarveenChadha & other friends from @SarvamAI Here's to more AI in 🇮🇳

~9 months ago @aakrit & @miten were looking for someone to help them out with their angel investments I was a college student not worried about full-time opportunities & w/ no immediate plans for higher studies As fate would've it, I applied for the position & somehow got in

A career growth hack for young professionals - the most interesting opportunities don't appear on job boards/career portals In an increasingly connected, remote-first world, opportunities can be manufactured Your dream job could just be a cold email, Github Repo, Blog Post away

Satya Nadella is to #CloudComputing & Enterprise Software what Steve Jobs was to Personal Computing & Smartphones, IMHO He helped Microsoft thrive in the cloud & mobile-first world in '14 And today Microsoft is a leader as we move to a world of distributed computing #respect

"What if X builds that?", "Why hasn't X built/building that?" are 2 common questions founders field from investors. I've been fortunate enough to be on both sides of the table & here's a scratchpad of thoughts on the skeleton of a few better answers I've given/heard 👇

💜 @sliceit_ UPI product - 95% of my P2M UPI transactions this month have been on this Would love to hear your experience if you're a user Follow @177pc for more takes on #product & #business

#productmanagement is taking the accountability for building the right thing Engineers are the most expensive & valuable resource for a tech company PMs determine how & where that resource is applied to help the company get an ROI #startups

This is one of the reasons why I believe the future of software is open - #OSS for the win 💪🏼

Was listening to @kunalb11 & @SharmaShradha this afternoon. Absolutely loved it. Have made some notes sharing the key takeaways for those pressed for time. Core Argument: obsess over/chase/compare skills, success & money will always follow. Link: piped.video/watch?v=12AOcV9l…

The most incredible Indian tech story is the underlying infrastructure that enables UPI IMHO India has the only "public" platform in the world w/ a billion+ users Here's a deep dive on the same from ~12 months back bizit.substack.com/p/8-reima…

Very elated to see @dhiwise cross the 100K users on the platform milestone w/ over 3M screens processed @Vishalvirani91 was a Gen AI founder well before it was mainstream & has been relentless in his execution To put these numbers in context, we went from 2M to 3M screens processed in 83 days And we doubled the no of users on the platform from Mar '23 And this is just the start - expect to see a lot more cool stuff coming out from Vishal & team 🚀🚀

Notes from @satyan's masterclass on Product/Growth at TPS by @TheProductfolks, @Inc42

This is an incredibly important point by @martin_casado - something that very few people understand/appreciate. In traditional software development, infra & application logic have distinct roles. Infra provides the underlying resources (such as servers, databases & networking), while application logic governs behavior, validation & correctness. Traditionally, application logic is explicitly coded by developers & sits on top of infra. This architecture ensures predictable behavior, making debugging & scaling straightforward. With LLMs, this boundary blurs. LLMs can process data, infer context & generate responses dynamically, often w/o predefined logic. The logic isn't hard-coded but emerges from the interaction between the model's architecture, training data & inputs. This flexibility allows LLMs to take on roles that traditionally required complex, explicitly defined logic. However, it introduces uncertainty because the logic isn't always predictable or transparent. Reconciling this w/ traditional software is an interesting thought experiment - some thoughts below: 1/ Validation & Correctness: Traditional software relies on strict validation rules, ensuring expected behavior. LLMs generate responses based on training data & user inputs, which can lead to unexpected or incorrect results. Bridging this gap requires additional validation layers, often w/ human oversight or secondary systems for validation. 2/ Scalability & Performance: LLMs demand significant computational resources, affecting scalability. Traditional software might scale by adding more servers or optimizing code, but LLMs require specialized infrastructure (GPUs). Ensuring consistent performance w/ growing demand is a unique challenge for LLM-based applications. 3/ Transparency & Explainability: Traditional software's logic is explicit, making it easier to understand & debug. LLMs operate in a "black box", posing challenges to explainability. Bridging this gap requires techniques to increase transparency, like model interpretability & explainability tools. Would love to learn from others.

A lot of people think of #OpenSource as just a business model At @scaletogether, we believe OSS is more than that - it's how you build software @MDDushyant & I tagged teamed a bit to prepare the following landscape comprising of some popular OSS #startups We ❤️ OSS companies

A great proxy to understand how Venture Capital works is to study Google's M&A Each of Android & YouTube alone is worth way more than everything they ever wasted money on in terms of bad acquisitions or resource allocation #venturecapital #startups #funding

Early-stage #startups usually have a flat hierarchy But as they set up processes & begin to scale, hierarchy kicks in It's when it becomes important to scale w/ processes but w/o bureaucracy Amazon, Netflix are 2 great examples in #tech on how to do it consistently

Information, capital, force & production decouple from geography: AI, crypto, drones, robots Governance is next → code When the ledger rules, sovereignty migrates from land to logic

Amidst all the memes, admiration & chatter around #SquidGame, we tend to overlook this astonishing stat: "@netflix will spend $17B+ on content this year" To put some perspective, Apple & Samsung spent $18.75B on R&D last fiscal year The focus is paying off big time! #business

Next-generation of interesting products on top of @NotionHQ - 1) potion.so/ - custom website on Notion 2) notiontweet.app/ - Notion docs into tweets w/ analytics 3) float.so/ - Notion docs into online courses Expecting more w/ @NotionAPI

Every company around us is becoming a technology company & we're seeing new software stacks emerge to support them in this transition while existing ones are being rewritten And we believe India is well-positioned to play a huge role in this once-in-a-lifetime transformation

Agents shouldn’t break like cheap pens 🖊️ Pumped to see @composio bag $25M Series A led by friends at @lightspeedvp + a dream bench of operators to ship the "skill layer" where agents learn-on-the-job The current traction numbers are 🔥- With 100k devs, 10M+ daily calls, it already feels like Stripe‑for‑AI Congrats @GanatraSoham, @KaranVaidya6, @sujaychoubey, @jiteshluthra & the rest of the team Proud to have played a part during my @scaletogether days Let's make agents hustle harder than interns on deadline 🚀💪

Heard @gauravmunjal live for the first time during @YourStoryCo's Techsparks 2020 last week & these are some of the notes from his candid conversation with @SharmaShradha. While he was very honest, he still didn't share the list of the books he recommends for aspiring founders.

I was looking at the evolution of email & found it pretty interesting. Quick thoughts - It's the equivalent of the Swiss Army Knife for our digital activities - from personal communication to collaboration, file storage to marketing channel, it does all of it & much, much more.

h/t @realmeson10 & @last9io team for putting together a gem: last9.io/9/ An absolute must-read for any technologist building products And I can tell you from experience: @realmeson10 is among the most customer-obsessed engineering leader there is #tech #developer

💯Keeping some allocation for angels is an underrated hack of building long-term leverage in your 0-1/1-10 journey They often punch way above their ownership in their company & help w/ introductions to experts, customers, potential hires, future investors, etc + @aakrit @miten

Heard @manasjsaloi is killing it as the unpaid intern everywhere I respect the hustle. Following his lead - I'm doubling down as a sales & customer support rep for the founders I like & as a technical intern for the founders who let me be Why limit yourself to one unpaid gig when you can have 12? 📈

Skills people put in resumes that are table-stakes - 1/ In 2010s - MS Excel 2/ In 2020s - Notion @NotionHQ's breadth & depth of adoption in the #startup land is amazing Consider mastering the tool if you aspire to work at startups 🚀 @akothari @ivanhzhao

🤔 What does GTM actually mean? It's 3 things IMHO - 1/ Help customers understand the product (marketing/positioning/product marketing) 2/ Get customers to use the product (figuring out distribution) 3/ Get customers to actually pay for the product (sales) #startups #product

While looking at a startup, had to revisit my notes on the Industry's shifts through platforms & biz models & rediscovered a timeless gem by @miten, @RavBhatia, @therealjpk on "Use Case". Brilliant insights on how media companies think & deal with aggregators/social platforms.

A very effective pitch while raising an angel cheque (from my limited experience) - 1/ This is what we're building 2/ This is why we're building 3/ This is who we are 4/ Here's the market we're in 5/ Here's the product that we've built 6/ Here's a product demo

@NotionHQ's product team is on 🔥🔥 It's is slowly becoming an indispensable piece of the personal productivity & professional tech stack It's the new command center for quite a few personas - marketers, data engineers & developers All-in-one workspace, indeed #startups

LinkedIn is on track to do ~$12B/yr in '21 w/ LinkedIn ads making up a third of it Just to give some context - Adobe did $12B in revenues in '20 Adobe's market cap is ~$300B while LinkedIn was acquired for $26B 5 years ago Microsoft's M&A in the past decade are 💯

We ❤️ Developers at @scaletogether We love making them more efficient 🧑‍💻 And this theme represents ~50% of our current portfolio We're also building 🛠️ something to help many more such builders at scale We like the following 2 logos - what's your pick?👇

"Being talented doesn't guarantee success. Being dependable does" ~ It's a tough lesson to learn but a very important one

"What do Slack, Notion, Heroku & Figma have in common?" They're all freemium/free-trial, bottoms-up businesses where you're running multiple business units in parallel Typically - 1/ indie users 2/ teams that onboard via self-serve 3/ large enterprises #startups #business

🤯 Amazon's ad business is at ~$32B run-rate growing ~50% YoY For a perspective, YouTube is ~$29B growing 42% YoY And Netflix is ~$30B growing 17% YoY Amazon's side hustles, like its ad business, are as large as its peers in big tech #business #businessnews

The AI firehose is deafening 🗓️ Starting a new weekly series: decoding the week’s AI moves - and what they really mean for AI in India 🇮🇳 Will try to do this every Friday This week crystallized 3 meta-themes: (1) Compute pluralism - OpenAI signed for Google TPUs while slashing o3 prices by 80%. (2) Agentic tailwinds - from Mattel's AI toys to YC's demo-day darlings, workflow-first agents are everywhere & represent the largest opportunity today. (3) 🇮🇳 AI - capital, talent, policy all surging. Details ↓ 🧵

I've been discussing the following theory w/ folks for some time: Meta's OSS Thesis with LLMs Beyond a certain quality of LLMs, their specificity matters more than their general-purpose reasoning abilities, as LLMs will be used as interaction engines for knowledge embedded inside a backend comprising Vector Databases (VectorDBs), Knowledge Graphs (KGs) & Relational Databases. The path to this level of specificity isn't a scaling game; it's likely a distillation game. A smaller & open model will likely be more effective in this context, as costs, privacy & latency are critical for most enterprise deployment use cases. For OpenAI and Anthropic, exposing their capabilities as APIs for enterprises to build on top of is beneficial. It allows them to generate revenue to fund their next training runs—pre- & post-training. However, this approach requires GPU clusters to be set up for inference, adding significant overhead. Meta, coming from behind in the LLM race, can leverage open sourcing as a strategic move. By open-sourcing Llama, Meta can benefit from the community's contributions to distillation & optimization, ultimately achieving a smaller model w/ quality equivalent to frontier models. Moreover, by open-sourcing, Meta doesn't need to worry about GPU provisioning for inference, allowing them to allocate GPU clusters to continue to chase scaling laws in the AGI race. The AGI race is particularly focused on consumer-centric use cases—an all-purpose AI that can handle everything from law to healthcare, education & personal assistance. Unlike their consumer-focused counterpart in the AGI race, Google, Meta doesn't have an in-house ASIC as mature as Google's TPUs. By offloading inference to the broader community, Meta can circumvent some of the infrastructure limitations they face compared to Google. Open sourcing also serves as a strategic maneuver to cannibalize the enterprise revenue streams of other foundational model labs, like OpenAI & Anthropic, while pursuing their own AGI ambitions. By giving away high-quality models, Meta potentially disrupts competitors who rely on closed models for monetization. In this way, Meta plays both the long game for AGI development & the short game of disrupting the enterprise market for other players in the LLM space.

@Rahul_J_Mathur - brilliant point The following are my favorite ones - 1) Netflix's signature "TaDum" 2) WhatsApp Web's Notifications 3) Command + Shift + Delete in Mac 4) Britannia's Ting Ting Titing jingle

"How is a startup different from a business?" My honest attempt at decoding the above 👇🏻

Public ☁️ is on 🔥 It's a neutralizing force in a global inflationary environment 1/ 🤯AWS is on track for $64B+ ARR in '21, growing 39% YoY To put some perspective, if it were a separate company, it'd have been 51st in Fortune 500 just behind @Disney & @LockheedMartin

Some popular tools that 0-1 early-stage startups use (across B2B & B2C, ~40 conversations) - 1/ ☁️Infrastructure - AWS/GCP 2/ Emails - G-Suite 3/ Storage & Document Sharing - Dropbox/G-Drive 4/ Async Chats - Slack 5/ Async Video - Loom 6/ Project Management/Kanban - Notion

We worked together for what is still the most memorable ~8 months for me In that period, we did over 15 deals & also saw a couple of markups for the portfolio companies I also got a chance to invest in a few companies in personal capacity - something I never thought I'd do

OSS is on the rise - we're seeing many #OpenSource alternates to many popular applications (Segment, Airtable, PostHog, Retool) But most OSS products need dev effort to set up, maintain & scale I suspect we'd soon see a #lowcode tool that abstracts away some of the complexity

💯 Underrated productivity hack for folks who share their no w/ folks on social media for WhatsApp chats Instead of typing down your no in the chat, share this link - wa.me/91<insert your 10-digit number> "91" is country code - modify it as per your location

🔥🔥 Composable, serverless AI at its best Huge leap by friends at @composio - fully managed MCP servers for @AnthropicAI Claude, @cursor_ai, @windsurf_ai & more Zero server overhead, integrated auth & frictionless function calling Has to be the blueprint for serverless AI With Composio's managed MCP, you get: ✅ Integrated & seamless auth ✅ Zero infra overhead setup ✅ Built-in & automatic scaling + HA AI infrastructure is now truly a serverless, composable service - isn't just about ops reduction but a new paradigm for building resilient, scalable AI apps

@venkyHQ: any company that has a transaction element will soon find a way to get into credit (Fintech) Any company that needs some degree of user education through videos/blogs/tutorials to drive adoption will get into content (EdTech) ^^ This is how I connect the dots

Was revisiting my notes on @kunalb11's now-famous Delta 4 theory. It's been compiled from multiple sources but not sure if this is the latest version. If anyone has any other documentation of the same, please share. Also, open to any feedback on this.

Did this quick exercise w/ an early-stage #SaaS founder From personal experience: Writing a #product positioning doc helps in - 1/ Clearer thinking 2/ Sharper articulation Would love to get some feedback from #startup folks to learn some best practices 🙏

An honour to host @vishalmisra (Vice Dean of AI @CUSEAS & the accidental inventor of RAG) for an exclusive session w/ Bangalore's top AI builders. Vishal was predicting GPT's capabilities & limitations much before anyone else. We skipped the hype & dove deep into what's next🧵 The consensus? The initial "wow" of LLMs is over. The real challenge is solving the messy, unglamorous "Day 2" problems of getting AI to work in the enterprise. Three critical themes emerged: 1/ Agent Security is a ticking time bomb 💣: As agents multiply, the attack surface explodes. Old paradigms for access control and data boundaries won't work. How do we govern thousands of autonomous agents securely? 2/ The MLOps & DevOps of the Agentic Era: Building one agent is a fun experiment. Deploying, monitoring, testing & managing a fleet of thousands is the real engineering beast. This is the next wave of DevOps/MLOps. 3/ New GUIs for human <> AI collaboration: To unlock the full potential of AI, software itself must evolve. The user interface needs to intelligently package context & state management into the product experience. Huge thanks to Dr. Vishal for his incredible foresight & the founders who joined for the sharp, honest debate. I'm currently partnering closely w/ some stellar teams pushing the envelope in AI. If you're also working on groundbreaking AI products, infrastructure, or research - I’d love to connect & exchange ideas. Thanks to @RTinkslinger & our good friends at @DeVC_Global for helping us bring the AI community together in their office. @aakrit @MShadagopan @ojasvi_yadav

🤯 AWS #cloud is on a tear 🔥 2020 Revenues: ~$45.35B 2021 Revenues: ~$62.2B That's a ~37% YoY growth in revenues Q4 '21 Revenue: $17.8B, growing 40% YoY And all of this at a 30% operating margin It'd have been 53rd on F500 as a standalone biz, easily worth >$1T

Cashbacks & discounts are an optimization problem in #product The "what" that needs to be optimized changes depending upon where you are in the product's lifecycle It's an extremely deep & complex problem that is solved in a data-driven way Unpacking some thoughts 👇

Given that we were sector & stage agnostic, it was a wonderful exposure to the intrinsics of different industries & operating models The common vision was to help missionary founders build the future & that kept us grounded

Increasing sales & marketing spend as a % of revenue & COGS will push SaaS from sales-led to product-led @Toplynehq is enabling & assisting companies to navigate this shift 🚀 Watching them execute from the sidelines has been a privilege 🙏 More: techcrunch.com/2022/05/03/ba…

@aakrit @miten - grateful for the trust & opportunity to shape this engagement as we went along 🙏 To the founders who trusted us: it was an honor & a privilege to walk alongside you & watch you scale - being on this side was interesting because of you guys 🙏

Prediction: #WhatsApp could soon become the channel for customer support for B2C apps The infra needed is already in place - WhatsApp Business APIs & User Phone Numbers CRED, MMT & some banks anyways use it for updates How hard is it to layer on customer support? #product

📌 What is MCP? The USB-C for AI Content Lately, there's a lot of chatter around MCP, its utility & confusion around fully understanding the potential & how it evolves from hereon Having spent a lot of time thinking on this topic, here's a sincere attempt at distilling my thoughts on the topic👇 MCP aims to simplify how LLMs access & utilize external information, much like how USB-C provided a universal standard for data & power It's an open protocol standardizing the way apps provide context to LLMs/AI systems - Open Protocol: Designed to be a community-driven standard fostering interoperability & avoiding vendor lock-in - Context: LLMs are context unaware instruction followers (h/t @NirantK). For them to be truly useful in real-world applications, they need to interact w/ current, relevant & often private data & this context could range from files on your local PC & data inside databases to info from APIs & content from web pages 📌So what is the big deal? Why is there so much interest? The real innovation & significance here lies in data isolation - Universal standard for data connection: Currently if you want to connect an LLM to an external data source, you often have to build a custom integration - time-consuming, complex & unscalable. MCP provides a single, unified protocol to connect AI systems with diverse data sources. Instead of building fragmented, point-to-point integrations, you build against a standard. - Simplified development: Imagine you want to build an AI assistant that summarizes docs from Google Drive, answer questions from a Postgres database & trigger actions in Slack. W/o MCP, you'd likely need to write custom code for each of these integrations. With MCP, you can potentially use pre-built "MCP servers" (connectors) for these systems or easily build your own knowing you're adhering to a common standard. This drastically simplifies development making it easier to scale AI apps across different data environments - Improves the quality & relevance of AI responses: By giving LLMs standardized access to more relevant context, MCP directly leads to better, more accurate & more useful AI responses - Open & community driven: Encourages collaboration, innovation & a wider ecosystem of tools & connectors 📌How did people do this before MCP? Why devise a new way? It's a different way of building AI applications & how they interact w/ data - Replaces Custom Integration Code: Instead of writing bespoke connectors for every data source, developers can leverage or build MCP servers reducing boilerplate code & dev time - Enhances existing tools: Products like @windsurf_ai, @cursor_ai, etc can integrate MCP to provide AI agents w/ better context about coding tasks. This could mean AI assistants that are more aware of your project structure, codebase & related documentation. - Complements Function Calling/Tool Use: Many LLMs already have "function calling" capabilities, allowing them to interact with external tools. MCP can standardize how these tools are connected & how data is exchanged, making tool use more robust & easier to manage across different systems. - Enables Agentic Systems: By providing a standardized way for agents to access & share context across different tools & datasets, MCP enables AI systems that can perform more complex, multi-step tasks & maintain context as they move b/w different environments. How does this evolve from hereon? 1/ Auth & Permissioning in MCP: Current Problem MCP is designed to connect LLMs to diverse data sources, many of which will be sensitive or require specific permissions. We can't just have open access to everything. We need robust mechanisms to ensure: - Identity Verification: Knowing who is accessing data (user, application, agent) - Authorization: Controlling what they are allowed to access & do - Secure Data Handling: Protecting data in transit & at rest within the MCP ecosystem Possible solution ideas (tech) - MCP identity layer: Define a standardized way for MCP clients (applications, agents) and MCP servers (data source connectors) to authenticate each other. This could build upon existing standards like OAuth 2.0 or similar, adapted for the MCP context. - Fine-Grained Permission Models: MCP needs to support granular permissioning at the data source level. This means MCP servers would need to translate MCP authorization requests into the native permission models of the underlying data systems (database roles, file system ACLs, API keys). - Policy Enforcement Points: MCP infrastructure would need policy enforcement points (likely within MCP servers) to evaluate access requests based on identity & permissions before granting data access. - End-to-End Encryption: Mandatory encryption for data in transit within the MCP network to protect confidentiality. Potentially support options for data-at-rest encryption within MCP servers as well. - Audit Logging: Comprehensive audit logs of all data access requests & permission decisions for security monitoring & compliance. Possible productized ideas - Simplified credential management: Maybe a central "MCP Credential Wallet" for users to manage their data source connections & permissions. - Visual Permissioning Interfaces: For devs & admins, GUIs to easily define & manage permissions for different users & applications accessing data via MCP. 2/ Routing & Context precision in MCP: Current Problem When an AI agent makes an MCP request, how does it know which MCP server to query? And how do we ensure it gets the most relevant context, not just any context? We'd need mechanisms for: - Service Discovery: MCP clients need to discover available MCP servers that can provide the context they need ("find me an MCP server that can access Google Drive documents") - Contextual Routing: Directing requests to the right MCP server based on the type of data being requested & potentially even the content of the query - Precision & Relevance: Ensuring that the context returned by MCP servers is highly relevant to the AI agent's query, minimizing noise & maximizing signal Possible solution ideas (tech) - MCP Registry/Directory: A central registry where MCP servers can register their capabilities (data sources they connect to, types of data they can provide etc). MCP clients can query this registry to discover suitable servers. - Semantic Routing: Beyond simple keyword-based routing to semantic understanding of context requests. Using metadata & potentially even lightweight ML models to route requests to servers that are semantically aligned with the query's intent. - Contextual Metadata Standards: Define standardized metadata schemas for describing the data provided by MCP servers. This metadata can be used for both service discovery & more precise context retrieval. Thinking of it as "tagging" data sources w/ semantic labels. - Query Refinement & Filtering: MCP clients could provide more structured or refined queries to MCP servers, allowing servers to return more precise context. MCP servers could also offer filtering & ranking mechanisms to prioritize the most relevant context. - Advanced querying: In more complex scenarios, MCP could support federated queries where a single client request is routed to multiple MCP servers & the results are aggregated & ranked for relevance. Possible productized ideas - Intelligent Context Suggestion: For developers, tools that help them discover & select the right MCP servers & define precise context queries. Maybe an "MCP Context Explorer" GUI. - "Smart Connectors" w/ auto-routing: Pre-built connectors that automatically handle service discovery & routing, abstracting away the complexity for less technical users. - Relevance Feedback Loops: Mechanisms for AI agents (+ users) to provide feedback on the relevance of context received, allowing the routing & precision mechanisms to learn & improve over time. These are the things I could come up w/ & I'm sure I'd have missed a lot of things here - please do share more ideas & feedback🙏

The Cash on Delivery (CoD) Payment Option was the OG BNPL model designed for India by Flipkart cc: @RachitaKumar7

Today is the 8th anniversary of Satya Nadella becoming the CEO of Microsoft In these 8 years, he's ~8x-ed the stock There are many amazing lessons to be learned from his tenure for all tech builders Sharing some of my favorites below 👇🏻 1/

A friend in the US mentioned - 1/ Lot of people left the valley from 2001-03 as very few #startups made through the Dot-com burst 2/ B2B was called "Back to Banking"; B2C ~ "Back to Consulting" He asked something I'll ask the community: "How's the expected scene in India now?"

All of #education for me is about: reading critically, reasoning analytically and thinking pragmatically

2020 should've been the year that ended prediction pieces but predictions are less about accuracy & more about reasoning. In that spirit, I sat down & chronicled my thoughts around some of the macrotrends that I think 2021 will be about 👇 A 🧵- bizit.substack.com/p/11-what…

"What are the moats in b2b SaaS?" Follow @177pc for more content on #SaaS & #products

#fintech & financial services have historically been glocal businesses #crypto /Web3 'fixes this' & turns it into a global industry h/t to @TheFinaveGuy & @Finave_in team for discussing this yesterday that helped me crystallize this But what do I mean by glocal & global here?

@NotionHQ templates are the best place for a beginner to learn about how tools & processes like CRM (Sales)/PRD (Product Management)/Project Management/Marketing work in the real world If you're not tinkering w/ them, you're not using Notion right

In product-led growth/PLG/bottoms-up SaaS, the atomic unit of sales funnels moves from a few macro accounts to many micro accounts This is where a tool like @RevenueHero helps sales reps prioritize the right accounts at the right time #startups

There are a lot of nuances that need to be considered for a proper GPT wrapper discussion, @vaibhavbetter (After 25+ conversations w/ app layer builders) Here's my best attempt to keep it crisp w/o compromising on nuance👇 (1) ChatGPT was launched on 30th Nov '22, - Until then, OpenAI's products were only developer facing APIs - The world, then, only had one frontier quality model available as an API - We already had products w/ real usage & revenues built on top of OpenAI's APIs: Copy.ai, Github Copilot, Jasper.ai - It was the first time the vendor providing those APIs moved up the stack into the app layer (2) Back then, application companies came in 2 shapes: either GPT-wrappers (Jasper.ai) or vertically integrated (Character.ai) - Both types of companies had great traction: revenues, usage & were founded before ChatGPT/GPT3.5 came - If you were a GPT wrapper, a lot of the value you were creating was around abstracting away the complexity of prompt engineering for the end user via your UI - Being a vertically integrated AI company needed massive capital investments to win compute/talent - It massively helped if you wrote the OG transformers paper &/or worked/interned at OpenAI/Google Deepmind - Most app companies didn't have much choice outside of being a wrapper around the frontier model API & winning on distribution/end-user brand recall - Hence, it did seem like incumbents are likely to run away w/ the market as they had both: deep pockets for talent/compute & brand/distribution (3) Things remained like this for a while until Meta came up Llama in Feb '23 - Around this time (end of Q1 '23), we also had Claude 1 & Vicuna also launch - Meanwhile OpenAI kept itself in the lead w/ GPT4 but the open-LLMs were comparable enough w/ GPT3.5 for applied AI use-cases - This was the first time application developers had access to a wider selection of models to choose from & wrapper companies became viable from a market timing perspective (very easy to call that in retrospect but I was in the middle of it & I know how hard/scary it was to be an investor) (4) Important to also note that then most of the initial application use-cases were content generation/analysis - Both of these categories of products lent themselves well to larger horizontal players (ChatGPT itself, incumbents like Notion/Coda/Airtable) vs niche startups focused on point use-cases - This isn't the fault of startups: if you had limited ability to fundraise, you could only tackle niche problems at the onset - This is the time when most people were solving for some version of marketing/sales copy generation, enterprise search or some LLM-induced sprinkle on the UI of an existing large horizontal player at the app layer (5) But things have changed for the better now - A lot of surrounding innovation in tooling, capabilities & infra along w/ a wide range of model selection available for almost every imaginable use-case is a major unlock for folks building at the app layer - Being a wrapper in this world doesn't (just) mean abstracting prompt engineering for the end-user - There do remain advantages of being a vertically integrated app company in some categories That said, my personal belief is that great companies win or lose on the field: there's only so much analysis of market dynamics (app layer vs infra layer) & product strategy (wrapper vs not) can do Copying some publicly announced portfolio companies/founders we've learned this from/with: @Tunehq_ai, @metaformsai, @sumanthr, @dhiwise, @PresentationsHQ

This breakdown of the main parts of the job of a Product Manager & the art of product management by @brian_armstrong is one of the finest I've read 💯 recommended for PMs & early-stage founders #startups blog.coinbase.com/a-letter-t…

1/n An interesting metric in the OTT space is "cost/first stream". Essentially, cost = marketing + production expenses. This when divided by the total number of people who watched the show/movie after signing up gives the metric. Meaning: The lower the number the better.

(At least) In India, we've always paid more for community/personalized services than content in #edtech Tuition teacher fees (services) = ~10-12x of NCERT Books (content) Digitizing both sides, we'd get Cohort-based Courses (CBCs) on LHS & MOOCs on RHS CBCs could be huge in 🇮🇳

Hiring managers/founders of early-stage companies tend to underestimate the impact a detailed JD/right on-the-job expectation setting can have in competitive hiring markets where everyone is vying for the same mind & effort share of a select no of individuals from the talent pool

A lot has been said about this topic but this is just half the story IMHO People clicking on more ads makes some of the very crucial functionality (search, social) of the internet FREE And it does help people & the society - which is what smart minds are supposed to do, right?

🚨@composio is designed to make AI agents useful 1⃣Even the smartest AGI-level LLMs won't know who things like your customer is & what's their most recent order - a business would need a reliable & secure way of connecting LLMs w/ software systems 2⃣Connecting, managing & optimizing interactions b/w AI agents & external tools requires a comprehensive toolset & framework & Composio acts as the central hub Check out composio.dev/ for more @KaranVaidya6 @GanatraSoham @composio

I've followed @sachdh's love for RL for 2+ years at this point & it's so great to see him help ship this This is interesting for a whole host of reasons but in particular because "scope" is beating "scale" Enterprise AI is tilting from “giant‑generalist” AGI aspirations toward Application‑Specific Intelligence (ASI): compact, open‑weight models fine‑tuned on proprietary data, running on dozens - not tens‑of‑thousands-of GPUs Some broad-based product development strategies that I am seeing/expecting from this trend: (1) Anchor on proprietary data gravity: Don't train but post-train where proprietary task/process data lives (2) Ship a ‘Model Lifecycle OS’: ingestion → fine‑tune → eval → monitor → iterate in one pane of glass; enable rapid experimentation w/ policies (GRPO/PPO/DPO), rewards & environments (3) Design for swapability: Abstract core logic from a single model vendor (4) Exploit agent telemetry: Instrument every prompt/action pair for future RL & learning loops Value creation in Enterprise AI is clearly shifting from model scale to application specificity @OpenAI launching an open-weights model & @miramurati's company announcing a massive fundraise around the same timelines is already suggesting the shift is underway

Umm, this is how I see it - 1/ Trillions were flushed into the economy 2/ This eventually reaches businesses or banks - their earnings/profits start looking better 3/ FDs were flat while Real Estate was down 4/ Public markets suddenly look very attractive & hence the activity

"Client vs Customer" They're used interchangeably but shouldn't as they mean very different things Lawyers/Consultants/Bankers/IT Services have clients (services industries) Product companies have customers How you use them tells a lot about your vision & aspiration!

Helpful #productmanagement mental model While building a B2B #product, think "Whose job am I going to enhance using my product?" For a B2C product, think "What job is your end-user hiring/paying your product for?" #startups

A marketing masterclass from @warikoo - 1) Think of consumers as being at the intersection of time, money & trust 2) Indians are time rich, money & trust poor 3) Design marketing campaigns for time rich, money poor & trust poor audience - aim to hit 2 on 3 for success

@malikgarv's post sparked a thought: Many edtechs are in the business of selling content When you target K-12 or test prep, the shelf-life of the content tends to be higher due to standardized curriculum It's difficult for upskilling related content as they tend to age quickly

.@dhiwise's new avatar allows you to deploy your React applications directly on Vercel & auto-identifies icons to set up actions like auth, validation, etc No wonder they're trending as the #1 product on Product Hunt 🔥 @Vishalvirani91 @rahul__shingala producthunt.com/posts/dhiwis…

Possibly the most underrated DevTools company is: $TEAM - ~$3B run rate this year - Bitbucket, Jira & Confluence have a self-serve experience (even at that scale) ^^ Proves that a company can be enterprise-focused w/o compromising on self-serve Lots to learn from for #startups

MS Excel is the biggest threat to B2B #SaaS #startups Some examples of how/where it's used - 1/ Sales & Marketing ~ Leads, CRM & Workflows 2/ HR ~ Payrolls 3/ Managers ~ Project Management PS - given the ubiquity of Excel, it's a non-exhaustive list w/ some general use-cases

📌It's difficult to tell the distinction b/w #SaaS & on-prem anymore A lot of SaaS cos today are separating their application's control plane from the #data plane Vendors own the control plane (just like typical SaaS) while the data plane is deployed in the customer's cloud/VPC

Studying the differences NSE/BSE & Crypto Exchanges is a nice way to appreciate the nuances of Web3/Metaverse & the opportunity. Crypto Exchanges - 1/ are always on 2/ can have participants from anywhere in the globe 3/ allow you to own & trade digital assets #cryptocurrecy

Some of the biggest businesses in India will be transactional & not ad-driven. ARPU for FB/YT ads in 🇮🇳 is <$5/user/yr. Macro-trend for 🇨🇳 & SEA too Hence, the interest for ad-driven companies to integrate transactions into their apps. Brilliant by @hchawlah & @NandanNilekani.

We probably need a Red Hat like business model that has enough consulting edge to win in the enterprise land grab & that has enough product chops to be able to build out a differentiated offering From my numerous conversations: - Enterprise AI adoption is hard & early w/ unproven ROI - Enterprises are currently preferring end-to-end offerings which often become bespoke implementations - Enterprises also need to understand a well-thought-through product roadmap to take a bet on something cc: @anshuman_01_ , @ChatNBX, @NimbleBoxAI

Freelance contributions to an early-stage #startup is a great way to see a fit You get to know - 1/ Founding team 2/ Products 3/ How do customers perceive it? 4/ Working culture Making a decision to join a company or not then is from a point of awareness

Given how loosely definitions & calculations go in the age of AI, @ChatGPTapp is likely a $150M+ ARR biz in India at this point At the time of QTing this, we had about ~98 people paying for ChatGPT/Claude on @anmolm_'s poll I'm guessing a good chunk of @CRED_club's userbase overlaps w/ this suggesting ChatGPT likely has 1M+ subs in India If @sama & team are seeing this, do consider enabling UPI auto-pay, supporting Indian payment gateways & having a slightly better customer support

.@aakrit's framework of going about angel investing for a beginner is pretty much what I followed & recommend to others It's a really smart, tactical way to get started As the ecosystem around #startups grows, we need more people to start angel investing

Friends at @composio have shipped something really cool here Composio SWE-Kit is designed to facilitate the development of Software Engineering (SWE) agents by leveraging Composio’s extensive tooling ecosystem There's a variety of tools on offer here including GitHub integration, repository indexing, file management & shell management, allowing devs to enhance agent capabilities Moreover, there's execution support across diverse environments like Docker, E2B & FlyIO ensuring secure & isolated operations tailored to specific project needs If you haven't checked them out yet, you should now: 1/ Website: composio.dev/swe-kit/ 2/ Docs: docs.composio.dev/swekit/ben… 3/ Do consider giving them an upvote if you like their work: producthunt.com/posts/swe-ki…

📌An underrated hack for aspiring PMs during their job search: for every PM, there are 8-10 engineers It's a thumb rule & directional, not a source of truth 1/ Check out the job listings of a company & find the no of engineering roles 2/ Check if there are PM roles pro-rata

Building & sharing in public is cool/cute But it's also important to understand that most of the very best pieces of work (investment memos, strategy documents, M&A proposals, proprietary software, sales pitches, etc) are still written, read & acted upon in private

Lovely guide by @dhiwise on leveraging @medusajs's flexible APIs & scalable back-end to build an e-commerce application Do give their platform a spin if you're a #Developer & let us know what you think @Vishalvirani91 & I would love to learn how we can improve

I'm humbled, privileged & super excited to be working w/ @mrgirish, @manav_garg, @avinashraghava, @shubg alongside @scaletogether team to support missionary founders in this journey🙏 It's time to put India on the map as a product nation #WeAreTogether #StrongerTogether

A few productivity shortcuts if you're a power user of web browsers - - doc.new or word.new to open up a new Google Docs or MS Word - slides.new or ppt.new for Slides/PPT - sheets.new/excel.new for Sheets/Excel

Spent some time on @HashiCorp's Q123 earnings & I felt it is a golden resource for early-stage DevTools/Infrastructure founders It's a typical modern-day Infrastructure SaaS company w/ an #OSS centric business model Tried to unpack some key learnings & share them in a 🧵👇🏻

Despite a host of alternates, @Calendly's moat is likely network effects Allowing recipients to book slots w/o conflicts by seeing their calendar on the same real estate is 🔥 IMHO Enabling that would need a critical mass of people to sync their calendars w/ the tool #startups

He does it for capital and in the internet age, capital isn't just financial. :)

Tired of wrestling with AI agent authentication?😩 @composio's AgentAuth makes it a breeze!💨 AgentAuth handles authentication for 250+ apps, so you can focus on building awesome AI agents. 🚀 1/ They support all the major authentication methods like OAuth and API keys. 2/ Plus, AgentAuth integrates seamlessly w/ top frameworks like LangChain & LLMs like OpenAI & Claude. Security is paramount! 🔐 AgentAuth comes w/ strong encryption & enterprise-grade features like SSO & RBAC Do check out their website for case studies & documentation to see how AgentAuth turbocharges your AI development👇

@warikoo beautifully explains ESOPs in his latest video. ESOP value, vesting, cliff & exercise period are things not understood even by graduates from the top univs in India today. My notes along with the video link for you to follow along 👇 tellmepc.com/Understanding-E…