LLM research, systems and compilers | ax + dspy in TS | agent engineering. aithy.dev axllm.dev graphjin.com
- Tweets 16,728
- Following 1,722
- Followers 5,213
- Likes 7,121
Pinned Tweet
we're on track to building the best dspy powered llm framework in the typescript world
4
5
83
22,351
ax has now reached 800+ stars. seems seems small in the scheme of things but it's how all things start. it's hard work, coding, evangelizing, helping folks use it. i'm convinced it's the best framework to build llm stuff with so i'm willing it into existence.
4
18
167
15,285
with all this talk around agent builders this is how simple it is to have a multi agent program in ax all driven by dspy signatures and optimizable.
9
8
153
12,509
dspy now works great in the browser. try the cool new dspy notebook with built in local in browser models. @DSPyOSS @lateinteraction you folks will like this.
7
18
139
8,147
this is why finance bro's wear those vests their offices are kept at the cooler than normal to keep folks active and alert
1
98
18,849
Hey H1B's I was one of you until I got a PR in Toronto now working on my own startup. I highly recommend Canada 🇨🇦 Ask me anything. #H1B #H4EAD #Greencard movnorth.com/story/
11
23
94
most schools are just holding facilties for kids, we so casually waste the best years of their life
1
3
83
6,067
typescript folks we got you covered github.com/ax-llm/ax
3
6
91
12,240
webllm running dspy in the browser. ax has come far i'm excited to build more, axflows, optimization, etc. models like gemma 3n make edge ai a very real thing.
2
3
73
4,271
It depends I built this thing which saves me hours. It auto compiles GraphQL to SQL. Supports deeply nested queries, inserts, updates, recursive queries, etc github.com/dosco/graphjin
6
6
69
why dspy usually wastes your time (and when it doesn't)
i could have done so much instead of build ax the typescript dspy library like riding my bike, reading books, watching anime or just taking a walk.
also now i have to respond to people who need help with the library so thats also time spent
but for building almost anything with an llm it's amazing i can't recommend it enough.
4
1
73
4,083
dspy is the only real way to building with smaller models which also include open source ones. it's insurance against that one day they decide to turn off access.
1
4
57
3,877
open source dosen't have to who cares about token costs when you can run your own private llm
1
49
5,605
My fist post on devto. REST vs GraphQL in 20201
dev.to/dosco/rest-vs-graphql… #GraphQL #Postgres #MySQL #Golang
7
48
every one of karpathy's slides is just a simple dspy program
Wow, @karpathy just dropped this gem at @ycombinator AI Startup School—Anatomy of @perplexity_ai and @cursor_ai
3
4
52
5,828
this is a system of llms (aka dspy program) that combines multiple smaller dspy programs to achieve bigger things.
1
3
49
3,552
the most group overworked in all of tech. i bet they're ai maxing like we're running out of tokens
2
1
48
5,047
when i was forced to develop for symbian i sent the ceo of nokia a 2 page email on why he should drop it and go with android he replied in short that smarter people where building symbian and it'll be the best thing ever soon
2
1
41
5,346
latest ax release has the new dspy optimizer gepa, mcp with oauth, and other fixes. github.com/ax-llm/ax/release…
2
6
45
15,978
everyone is out of ideas, except for dspy, the privacy preserving agent in dspy python is freaking cool we need more ideas like this
3
4
43
5,557
agents are just code with llm calls mixed in. using dspy makes those llm calls fit in with the rest of this code. we tried gui code builders before nothing beats just writing the code
Introducing AgentKit—build, deploy, and optimize agentic workflows.
💬 ChatKit: Embeddable, customizable chat UI
👷 Agent Builder: WYSIWYG workflow creator
🛤️ Guardrails: Safety screening for inputs/outputs
⚖️ Evals: Datasets, trace grading, auto-prompt optimization
3
6
44
5,919
often you can tell how based something is just off the vibes and i support these vibes. usa usa
41
6,103
we need a dspy signature standard and extend that so we can define whole programs with signatures
6
2
40
5,350
open source is the darwinian battle of code let them fork and build. more power to that!
1
3
36
11,014
I built Super Graph a GraphQL to SQL compiler in GO. I wrote it cause I did not want to deal with SQL. How would you write the query to fetch the data for this below GraphQL? Only 1 SQL query allowed. Also topics is an int array column on threads that is related to a tags table.
3
3
34
Super Graph compiles all GraphQL queries even nested inserts and updates into a single SQL query to keep things efficient. Sorry had a hard time exporting the image download and zoom to read.
6
1
34
Build a MongoDB like NoSQL JSON database on Postgres with one line of SQL.
CREATE TABLE jsonDB ( data jsonb );
1
4
36
I'm working on an automatic GraphQL to SQL compiler in Go. It does discover on your DB and generates a single efficient SQL query for even deeply nested GraphQL queries and mutation. github.com/dosco/graphjin
2
1
30
you need stuff like this for a production agents aka dspy programs aka systems of llms. ax supports all of this i literally built it to put dspy in production
Google Cloud Tech
2
1
35
2,423
the coolest thing about llm's is that they're just text in -> text out, adding a lite structure on top of input fields -> output fields exponentially increasing the ability to compose these together into systems of llms this is the key to why dspy matters.
1
5
34
2,945
llm-client the best way to build with llm's - api calling, cot reasoning, simple, sensible defaults and support for openai, cohere, claude, google, together, etc github.com/dosco/llm-client
1
3
29
3,901
ax is a framework built for creating agentic workflows based on the stanford dspy paper. at it's core is the concept of prompt signatures, everything is built on this even agents and teams or agents.
2
4
30
7,052
As a huge fan I can’t believe you signed onto this. Asking big brother to crush the little guys experimenting towards a better world. As for the others creds in CS != understanding of crypto
1
2
30
we have a exciting roadmap towards building the best llm framework in the typescript world
Yes @dosco’s Ax is the unofficial (semi-official?) DSPy in TS/JS.
2
3
25
8,867
all credit for the gepa implementation in ax goes to its author @monotykamary github.com/ax-llm/ax/pull/34…
Ax brings GEPA + GEPA‑Flow to TypeScript. You don’t chase a single score anymore - you evolve a Pareto frontier (e.g., accuracy and brevity and cost) and pick the trade‑off your product needs.
More complex code example for moderation with rationale discipline at the very end ↓
2
29
1,997
ai wrote all my code today 100% of it. i spoke to ai for hours about designs and ideas, i shipped multiple massive commits to ax and had had more time than ever to ride my bike, think, read and listen to music. panic, this is not normal
3
2
28
1,905
I'm excited to unveil 'Ax' the best framework to build powerful agents, agentic workflows, and anything else thats powered by LLMs. This is a pure open-source labour of love that evolved over a couple years.
axllm.dev
3
7
28
1,718
and now they're being brainwashed to learn useless sever ops instead of building their product and talking to customers
1
27
9,308
My kid aka DJ Nocturnal playing an amazing self created set on @sonic_pi cc @samaaron @simplyflipflops
3
3
24
on the dwarkesh pod karpathy ended a lot of model hype, he also said incontext learning and hence dspy is the right approach
26
1,639
dspy prompt signatures add structure to the text blob that is the prompt. you now reason about it in terms of inputs and outputs. on this abstraction you can build anything agents, tuning, composable prompts, systems of llms, multi-modal stuff, etc
1
3
25
12,008
checkout ax the dsp powered typescript agent building framework works great with anthropic claude. multi-modal, streaming parsers, dsp, rag tons of useful stuff. github.com/ax-llm/ax
1
4
23
4,016
these guys are building something really amazing, async coding is whole different ball game from vibe coding.
Meet the Jules Critic, a new feature that acts as Jules' built-in code reviewer. It evaluates the code Jules generates, flagging potential bugs and tightening the logic to ensure the final output is more robust and reliable.
We're rolling this out to users over the next few days. We're excited for you to try it!
More details at: jules.google/docs/changelog/… and a blog coming next week.
1
22
2,243
the guys discussing the ai in the robot at the end of the video is the most sf thing ever
1
23
3,578
With GraphJin just put the below code into a file and now you have a REST and a GraphQL API instantly.
GraphJin automagically coverts GraphQL into a working API endpoint. github.com/dosco/graphjin
5
21
lots of alpha trapped in expeienced folks working in faang. successful founder is a numbers game so you'll find signalling and conformity at the same rate as in the general population
17
7,531
openai if you reading this checkout this one. github.com/ax-llm/ax
1
21
1,905
Excited to share "Minds" with you'll. A new way to build app backends and workflows entirely with AI (LLMs from OpenAI and Cohere). The AI can call your APIs, lookup in your database, etc. 🙏
github.com/dosco/minds
4
3
22
5,664
with so much excitement around dspy i just want to remind typescript folks you have ax it’s dspy for typescript and has optimizers like mipro and gepa github.com/ax-llm/ax
Have heard great things about DSPy plus GEPA, which is an even stronger prompt optimizer than miprov2 — repo and (fascinating) examples of generated prompts at github.com/gepa-ai/gepa and paper at arxiv.org/abs/2507.19457
3
3
23
2,791
Just got done with a fantastic trip to the Maker Faire 2015. Enjoy the robot dance #MakerFaire
2
9
19
having a single gpu is hardly compute rich yet open weights give me unrestricted local ai
2
19
1,362
after dspy, "society of minds" is the other seminal paper in the llm space. working on something open source related excited to share soon arxiv.org/abs/2305.17066
3
4
21
968
my mlops community podcast just dropped. was extremely cool to be invited. watch and learn DSPy, Ax, LLMs, and all things based
3
4
21
2,098
as a dspy user it's always fun to see a big model launch claim something that your system of llms could always do with smaller models
2
21
4,551