Projects
01
GoSFU
LiveA real-time voice infrastructure built from scratch in Go to understand what frameworks like LiveKit hide. It handles the full media pipeline end to end: WebRTC signaling and transport, Opus encode/decode, resampling, VAD, STT, LLM, and TTS, with a frame pacer, jitter buffer, and backpressure keeping speech smooth under a few hundred milliseconds of latency.
02
DeepRAG
LiveA bare-metal retrieval stack built by hand to learn why every layer exists. It traces the full evolution of search: inverted indexes and BM25, embeddings and semantic search, chunking and contextual retrieval, hybrid search with reciprocal rank fusion, cross-encoder reranking, multimodal CLIP search, and an evaluation harness to measure what actually improves recall.
