AI & ML Engineer
– Present
Garage (Gravitichain Technology Group Pvt. Ltd.)
Bengaluru, India
- Built the AI layer of NetworkChains — an agentic sales copilot (5-round tool loop; a 7-layer prompt stack cut worst-case input from ~96K to ~2K tokens/turn), a realtime call assistant (dual-stream 48→16 kHz PCM16, one Deepgram socket per source giving speaker attribution without a diarization model), a relationship graph and conversational image editing — over 6 purpose-scoped Qdrant collections using contextual retrieval, HyDE, reciprocal-rank fusion and LLM rerank.
- Built the AI layer of Garage (workspace + commerce) — a provider-agnostic assistant normalising OpenAI/Anthropic/Gemini tool schemas, video and PDF Q&A, and a crash-recoverable LiveKit note-taker with adaptive map-reduce summarization (single
gpt-4opass under a 15K-token budget, elsegpt-4o-minichunks merged bygpt-4o) plus transcript preprocessing cutting 20–40% of input tokens. - Turned OpenClaw, an open-source single-user CLI agent runtime, into a multi-tenant service across both products: 2 FastAPI orchestration backends, per-user/per-org agent assignment with proxy-level RBAC, a custom 265-tool TypeScript extension spanning 57 integrations and 36 OAuth providers (Fernet-encrypted), a 40-tool MCP server, and WebSocket JSON-RPC with an Ed25519 device handshake.
- Shipped supporting AI services — voice-cloned dubbing across 20 languages, highlight clip extraction, ad-video generation, and a Voxtral STT/TTS telephony plugin at ~20× lower cost than Whisper; deployed via Docker + GitHub Actions on GCP and DigitalOcean.