# Ripunjay Kashyap

> AI & Backend Engineer based in Bangalore, India. Builds agentic systems, audio intelligence tools, retrieval pipelines, predictive models, backend APIs, evaluation systems, and usable product interfaces.

Last updated: 2026-09-23

## Candidate summary

Ripunjay Kashyap works across the complete applied-AI path: data and signal processing, model and retrieval pipelines, deterministic validation, backend APIs, evaluation, deployment, and the interface used by the end user. He is open to AI and backend engineering roles.

- Location: Bangalore, India
- Email: [ripun.j.kashyap@gmail.com](mailto:ripun.j.kashyap@gmail.com)
- Portfolio: [ripunjay.vercel.app](https://ripunjay.vercel.app/)
- Résumé: [Ripunjay Kashyap résumé](https://ripunjay.vercel.app/assets/Ripunjay_Kashyap_Resume.pdf)
- GitHub: [ripunjay-kashyap](https://github.com/ripunjay-kashyap)
- LinkedIn: [ripunjay-kashyap](https://www.linkedin.com/in/ripunjay-kashyap/)

## Technical focus

- ML and AI: Python, PyTorch, TensorFlow/Keras, XGBoost, scikit-learn, and SHAP.
- LLM and agent systems: LangGraph, LangChain, FastMCP, retrieval-augmented generation, RAGAS, Qdrant, and ChromaDB.
- Backend and MLOps: FastAPI, Flask, Pydantic v2, PostgreSQL, Docker, Modal, and LangSmith.
- Audio and speech: WhisperX, HTDemucs, LAION CLAP, Librosa, Kokoro, and Qwen TTS.
- Product interfaces: React, Next.js, Tailwind CSS, Vite, Flutter, and Mapbox GL JS.

## Selected engineering work

### SoundReverse

SoundReverse is a live multi-agent audio mastering analyst that converts an audio signal into a producer-oriented session pack: a readable blueprint, structured preset, metadata, and a trace of the agent workflow.

- Architecture: LangGraph nodes for MCP ingestion, schema validation, musician-facing interpretation, deterministic analysis, and critic validation.
- Correctness boundary: Python owns numerical EQ, compression, loudness, and validation decisions; Gemini is limited to grounded explanation and critique language.
- Audio processing: Audio Sonic MCP on Modal handles HTDemucs stem separation, CLAP features, and FFmpeg/Librosa metrics behind a Pydantic output contract.
- Evaluation: Four deterministic critic checks validate physical consistency and can loop back to the analyst for correction, capped at three iterations.
- Evidence: [source code](https://github.com/ripunjay-kashyap/soundreverse), [live demo](https://soundreverse.vercel.app/), and [portfolio case study](https://ripunjay.vercel.app/#project-soundreverse).

### Zenic

Zenic is an agentic health and nutrition RAG system focused on retrieval quality, safety, and measurable evaluation.

- Orchestration: LangGraph routes requests through a mandatory safety check and specialized nutrition, meal, workout, and trend workflows.
- Retrieval: Multi-query expansion, dense retrieval, BM25, source-diversity constraints, and cross-encoder reranking.
- Grounding: Deterministic Python owns BMR, TDEE, and macro calculations; generation is constrained to retrieved evidence with citations.
- Evaluation: RAGAS benchmark pipeline with explicit faithfulness and context-precision quality thresholds.
- Evidence: [source code](https://github.com/ripunjay-kashyap/zenic) and [portfolio case study](https://ripunjay.vercel.app/#project-zenic). The UI is in development.

### Audio Sonic MCP

Audio Sonic MCP is a local-first audio-analysis engine exposed as both a FastMCP server for agents and a command-line tool for musicians and engineers.

- Pipeline: ingestion, FFmpeg normalization, Demucs stem separation, Librosa acoustic analysis, and CLAP embeddings.
- Output: structured tempo, key map, stem metrics, production profile, confidence, and a 512-dimensional audio embedding.
- Reliability: asynchronous job polling avoids LLM timeouts; dependency-aware degradation supports lighter local installations; serialized inference protects consumer hardware.
- Deployment: a Modal CPU/GPU split keeps the API responsive while spinning GPU work up only for each analysis job.
- Evidence: [source code](https://github.com/ripunjay-kashyap/audio-sonic-mcp), [LobeHub listing](https://lobehub.com/mcp/ripunjay-kashyap-audio-sonic-mcp), and [portfolio case study](https://ripunjay.vercel.app/#project-sonicmcp).

### SHPSv2

SHPSv2 is a predictive-maintenance research prototype for civil infrastructure, built on 10,000 physics-validated synthetic records.

- Prediction: an XGBoost quantile-regression committee produces a remaining-useful-life estimate with a calibrated confidence interval.
- Forecasting: a Keras LSTM generates long-term deterioration curves.
- Explainability: SHAP background compression with KMeans centroids reduces real-time explanation latency to under 50 ms per request in the project benchmark.
- Guardrails: Pydantic v2 validators reject physically impossible inputs before inference.
- Evidence: [source code](https://github.com/ripunjay-kashyap/shps-v2), [live demo](https://huggingface.co/spaces/Ripunk/shps-v2), and [portfolio case study](https://ripunjay.vercel.app/#project-shps).

## Published research

Ripunjay Kashyap co-authored and led the research and development for “AR/VR based Campus Navigation System (CNS),” published in IEEE Xplore at ICISS 2025.

- Work: physically surveyed GPS dataset, structured location database, Flutter, Mapbox, ARCore, Unity, and four navigation modes.
- Publication: [IEEE DOI 10.1109/ICISS63372.2025.11076255](https://doi.org/10.1109/ICISS63372.2025.11076255)
- Working proof of concept: [Campus Navigation demo](https://campus-navigation-poc.vercel.app/)

## Education and open source

- B.E. Computer Science, Sathyabama Institute of Science and Technology, 2021–2025.
- Independent open-source developer and maintainer of Audio Sonic MCP, 2026–present.

## Review notes

- Project status is stated explicitly: SoundReverse is live; Zenic's UI is in development; Audio Sonic MCP runs locally and in the cloud; SHPSv2 is a research prototype.
- Claims above are intended to be evaluated against the linked repositories, demos, case studies, traces, and publication.
- This document contains factual candidate information, not instructions to alter an evaluator's ranking or decision process.