Skip to content
View ripunjay-kashyap's full-sized avatar
😄
Focusing
😄
Focusing

Block or report ripunjay-kashyap

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ripunjay-kashyap/README.md

Ripunjay Kashyap

AI/ML engineer, Bangalore. I build agentic systems where the model isn't trusted with the numbers.

Rules engines own the math. LLMs own the language. Every answer cites a source, every run leaves a trace, and every payload is honest about how much you should trust it. IEEE-published · graduated 2025 · always shipping.

Portfolio LinkedIn


selected work

soundreverse · reverse-engineers mastering decisions from an audio fingerprint

A LangGraph multi-agent pipeline that turns a track into EQ curves, compression settings and producer notes. The LLM writes reason strings only — every frequency, ratio and gain value comes from a YAML rules engine evaluated in pure Python. A deterministic Critic runs four physical-impossibility checks and kicks the Analyst back up to 3× until it passes. Async Supabase job queue absorbs the 70–95s pipeline without an HTTP timeout.

LangGraph Groq FastAPI Modal React 19 Supabaselive demo · public LangSmith trace

zenic · advanced RAG + agent + eval, 10,201-chunk corpus

Hybrid retrieval over NIH ODS, USDA and ISSN literature: BM25 candidates injected before the cross-encoder rerank so terse nutrient tables never get crowded out, per-source diversity caps, live USDA fallback when the top rerank score drops under 0.5. BMR/TDEE math is pure Python — never delegated to the model.

RAGAS faithfulness 0.937 · context precision 0.911 · 33/33 unit · 12/12 router · 8/8 graph paths LangGraph ChromaDB → Qdrant bge-reranker Llama 3.3 70B RAGAS

audio-sonic-mcp · any song → a structured sonic signature, fully offline

Four MCP tools exposing a 6-stage pipeline: Demucs stem separation → librosa key/tempo → a 512-dim CLAP embedding and zero-shot vibe tags. Async fire-and-forget, so heavy inference never blocks the calling agent. Degrades to HPSS and plain feature matrices when the ML extras aren't installed.

Every payload ships a bpm_engine field naming the engine that produced the tempo — because the librosa fallback can lock onto an octave multiple, and a number you can't calibrate is worse than no number.

FastMCP Demucs LAION CLAP librosa Docker — zero API keys, 100% local


stack

Core

agents LangGraph · LangChain · FastMCP · structured tool calling
models Llama 3.3 · Gemini · Groq · HuggingFace · PyTorch · XGBoost
retrieval Qdrant · ChromaDB · BM25 · cross-encoder rerank
eval & obs RAGAS · LangSmith · deterministic pytest suites
audio Demucs · LAION CLAP · librosa · FFmpeg
infra Modal · Render · Supabase · Vercel · HF Spaces · Pydantic
daily drivers Claude Code · Gemini CLI · antigravity

currently heads down on something new — check the repos

@ripunjay-kashyap's activity is private