Third-year CS undergrad working on the unglamorous half of AI — the part where a model has to answer a real user, in real time, without falling over.
- 🎙️ Building — a real-time voice RAG assistant on LiveKit + Sarvam, targeting sub-second turn latency
- ⚡ Learning — TensorRT-LLM, Triton Inference Server, quantization trade-offs at serving time
- 🔎 Curious about — agent memory, retrieval that doesn't hallucinate, and where inference cost actually goes
- 💬 Ask me about — streaming STT/TTS pipelines, FAISS retrieval, FastAPI + WebSockets
- 📫 Open to — AI engineering internships, backend/SWE roles, open-source collaboration
|
Real-time conversational agent over a private knowledge base. Streaming audio in, grounded answers out. Pipeline — Silero VAD → Sarvam STT → FAISS retrieval → context injection → LLM → Sarvam TTS, streamed over LiveKit WebRTC.
|
Benchmarking what actually moves the needle on serving latency and throughput. Exploring — TensorRT-LLM engine builds, Triton model repositories, INT8/FP8 quantization, batching strategies, and the accuracy cost of each.
|
💡 Swap these for your two strongest repos and add a one-line result to each — "cut p95 latency from 2.4s to 780ms" beats any badge on this page.
| Languages |
|
| AI / ML |
|
| Backend |
|
| Infra |
|
Training the model is the demo. Serving it at 3 AM to a user who doesn't care that it's AI — that's the engineering.