Graphsight: project memory for coding agents
It turns a repo's GitHub history into a typed graph, then lets an agent add the team facts GitHub doesn't hold. A tracer shows what the agent retrieved and what it actually used.
- Memory an agent can correct. A changed fact ends the old one and keeps it as history. A retracted fact stops being recalled.
- A source on every fact. The model is told which facts came from GitHub and which an agent was told.
- Measured, with the misses published. Tasks are written before any run and graded by exact match. The eval write-up lists what failed as well as what passed.
Python LadybugDB LangGraph MCP FastAPI React
Code · PyPI · Site · Eval
Saadho: AI/ML interview prep, day by day
Give it a role, an interview date and your daily hours, and it builds the plan. Live in production.
- A plan, in order. 300+ vetted resources sequenced by prerequisite, with what to skip.
- Company guides. How 200+ companies interview, round by round, with sources.
- A voice practice room. A group discussion with four AI colleagues in English or Hinglish, then a scorecard that quotes your own words.
Next.js Supabase LLM agents Realtime voice
Live site
I'm working down the stack, from agents to how a model is actually served. These are the codebases I'm reading with the aim of contributing.
| Repository | What I'm reading it for |
|---|---|
| vllm-project/vllm | Continuous batching, PagedAttention, and the serving scheduler |
| ggml-org/llama.cpp | Quantization formats (GGUF) and CPU/GPU inference in plain C++ |
| sgl-project/sglang | Prefix caching with RadixAttention and structured-output serving |
| NVIDIA/TensorRT-LLM | Engine building and GPU-side optimization |
| ModelTC/LightLLM | A lightweight Python serving framework: async processes and dynamic batching |
| skyzh/tiny-llm | Building an inference engine from scratch, to learn the internals |
| modular/llm-inference-handbook | The reference I read alongside the code |
Notes and experiments: inference-engineering0to100.
| Project | What it is |
|---|---|
| SmartFlow | A reinforcement-learning traffic-signal controller. Cuts average waiting time about 55% against a fixed-time controller, and extends to a multi-agent 2x2 grid with emergency-vehicle preemption. SUMO PPO/DQN · Code |
| Bodhi | A voice-first AI mock interviewer: an adaptive pipeline, RAG over 10K+ chunks, and real-time proctoring. LangGraph pgvector YOLOv8 |
| PADHO | Agents with hierarchical orchestration and RAG, in pure Python with no framework. |
| venvctl | A single-binary Windows CLI that manages Python virtual environments across PowerShell, CMD and Git Bash. Go |
| ReAct-CLI | A small command-line implementation of the ReAct loop: reason, act, observe. |
- Built ML pipelines processing 10,000+ aerial images a day at Kanchan Drones.
- Built a real-time exam-proctoring system holding sub-200ms latency on GCP.
- Shipped a policy-aware RAG chatbot that answers in under a second.
- Taught an 8-week NLP course to 20+ students.



