Building open-source AI tools. Created https://xitter.cf/t.co/gz7B4NoFhQ (50K+ views on Reddit in 48 hrs). Ships weekly. https://xitter.cf/t.co/GlW5keAV0x
Pennsylvania, USA
Joined November 2015
- Tweets201
- Following534
- Followers131
- Likes236
I built PageBot — add an AI chatbot to any website with one line of code 🤖
<script src="pagebot.themenonlab.com/widg…"></script>
That's it. No accounts, no config, no training data.
Thread 🧵
Why BYOK?
Most chat widgets charge per-message. AI costs drop constantly — the vendor pockets the savings.
With BYOK, you pay your AI provider directly (~$0.01-0.05/chat). Every price cut benefits YOU.
Prahlad G Menon, PhD, PMP retweeted
AATS Artifical Intelligence Architect @themedcave gives a live demo on the use of AI for practicing surgeons during the #AATS2026 AI Practical Toolkit session. Head to Room E353A to hear from @BMBurtMD and @DrRaviGhanta.
Karpathy's autoresearch hardcoded the loop to ML training. I generalized it!
autoloop: point it at any file + give it a metric
→ agent runs N experiments, keeps what works
pip install autoloop-ai
→ github.com/menonpg/autoloop
@karpathy 🙏 opened a PR to add to related projects
Been tracking a thread here for a while: AI systems that improve themselves rather than waiting for humans to tune them.
Imbue’s Darwinian Evolver mutates code like organisms. menonlab-blog-production.up.…
Ouroboros rewrote its own agent logic while its creator slept.
menonlab-blog-production.up.…
Sentrux gives coding agents a real-time quality signal to self-correct against. menonlab-blog-production.up.…
A new paper published yesterday adds a layer none of those touched: optimizing the harness itself.
Meta-Harness: End-to-End Optimization of Model Harnesses
Yoonho Lee (arXiv:2603.28052) — out of Stanford, preprint published March 30, 2026. arxiv.org/abs/2603.28052
CrewAI: 15,000 lines
LangGraph: 50,000 lines
The thing I actually needed: 150 lines
I built litecrew — multi-agent orchestration without the framework tax.
🧵
What it does:
→ Sequential handoffs (A → B → C)
→ Parallel execution
→ Tool calling
→ Token tracking
What it doesn't (on purpose):
→ No hierarchies
→ No state machines
→ No YAML
→ No PhD required
Works with:
• OpenAI
• Anthropic
• Ollama (fully local, no API keys)
• Any OpenAI-compatible API
BYOK — your keys never leave your machine.
MIT licensed.
github.com/menonpg/litecrew
pip install litecrew
Submitted SoulSearch to the Chrome Web Store 🔍
Chrome extension with AI that actually remembers you -- memory stored in YOUR private Git repo, not some company's servers
- reads any webpage in context
- Claude / GPT / Gemini
- agent mode (click, type, scroll via text)
MIT, no subscription
Open source: github.com/menonpg/soulsearc…
soul.py v0.2.0 just dropped 🧠
New feature: the Modulizer
→ 50% token savings
→ zero infrastructure (no vector DB)
→ one command: soul modulize MEMORY.md
your agent reads an index first, then pulls only what it needs.
pip install --upgrade soul-agent
Works with any LLM - cloud or local! github.com/menonpg/soul.py
One man. $3,000. No biology degree.
Sequenced his dog's tumor DNA, fed it to ChatGPT + AlphaFold, designed a custom mRNA cancer vaccine from scratch.
Tumor shrank 75%.
The professor: "If we can do this for a dog, why aren't we rolling this out to humans?"
The pipeline works.
The mRNA cancer vaccine pipeline that saved Rosie:
→ $3,000 tumor DNA sequencing
→ ChatGPT for research + strategy
→ AlphaFold for protein structure
→ pVACtools + NetMHCpan for neoantigen selection
→ UNSW RNA Institute for synthesis
→ 75% tumor reduction in one month