About
Signal, mostly, in one form or another.
I'm an AI/ML engineer based in Mumbai. I came to machine learning by way of electronics and telecommunication — antennas, breadboards, signal chains — at Vishwakarma Institute of Information Technology, and the habit never quite left: I still think of most of what I build as a signal moving through a system, whether that's a live phone call, a pulse trace on someone's wrist, or a player's guess at a secret a language model is trying to keep. These days that mostly means production speech pipelines at Emitrr, agentic systems that reason over documents, and the occasional paper on models that have to explain themselves.
- Top 3 of 500+ teams, Auto PCOS Open Hackathon (MisaHUB)
- 52nd on Hugging Face Hub's Leaderboard, July 2023
- Open-source contributions to ReX (explainable AI) and Gensim (NLP)
- AI / LLM & GenAI
- Transformer Architecture, LLMs, RAG, Agentic AI, Multi-Agent Systems (CrewAI, LangChain), Prompt Alignment, LLM-as-Judge Evaluation
- Speech & Audio
- STT (Deepgram), TTS (ElevenLabs), Streaming Audio, Endpoint Detection, PPG Signals
- ML / DL Frameworks
- Scikit-learn, TensorFlow, PyTorch, OpenCV, HuggingFace Transformers
- Languages & Backend
- Python, C++, SQL, Bash, Flask, FastAPI
- Data & Infra
- PostgreSQL, FAISS, Docker, Git, Weights & Biases, AWS, GCP, Linux
Research & publications
What I tend to optimize for
Different problem classes reward different design trades. These are the lenses I keep coming back to.
Speech + operational systems
The real differentiator is not the model in isolation — it is how the full streaming pipeline handles latency, turn boundaries, and failure states while the call is still live.
Live GitHub pulse
Public profile numbers fetched client-side from GitHub.