the long version: what I build at Northstarz
- Multi-agent recruiting assistant: Go + Google ADK (migrated from LangChain). It handles sourcing, screening, outreach, evaluation and scheduling, and has helped 30+ companies close 600+ positions.
- Candidate search: Gemini embeddings in Qdrant, hybrid dense + BM25 with Reciprocal Rank Fusion, ColBERT v2 re-ranking and LLM scoring. A 200-query live eval suite gates every change.
- Voice AI: a Gemini Live calling agent over Twilio, a WebSocket AI interviewer in English, Hindi and Kannada, and a WebRTC interview room with Go signalling and a self-hosted TURN server.
- AnthroSense: 13 BiLSTM trait models trained in PyTorch and served from Go via ONNX, at zero per-call LLM cost.
- Cost: incremental scoring, Flex-tier inference and model routing cut talent-matching cost from $0.24 to $0.005 per match.
| project | what it does | |
|---|---|---|
| 🧭 | job-application-agent | Finds matching jobs, tailors the resume per role (an auditor pass blocks factual drift), fills applications, drafts recruiter notes |
| 🕸️ | Agent Hub | Runs Claude, Codex and Gemini as one coding-agent pool: shared SQLite blackboard, deterministic routing, a "needs you" queue |
| 🎬 | Hermes AI Reels Studio | Topic in, published vertical reel out: Gemini script, Veo 3.1 clips, TTS, word-synced captions |
| 🐈⬛ | Void | The cat-shaped guide on my portfolio: ask it anything about my work |
| 🔤 | nanoGPT | GPT-2 from scratch in PyTorch, trained on Shakespeare |
| 🍲 | Novel Recipe Generation | GPT-2 / Mistral / LLaMA-3 on 118k recipes, validated by actually cooking them |
- fix(sharing): Keep the parent when trimming a revisions tree past MaxDepth in linagora/cozy-stack ★1,268
- fix(core): give a Concierge chat created before any login its model later in rome-os/rome ★893
also: Google ADK · LangChain · ONNX · Vertex AI · Gemini · Qdrant · ColBERT · NATS JetStream · WebRTC · Twilio
ML Engineer @ Wishfin: loan-eligibility scoring that lifted partner-bank conversions 70%
Research Assistant @ MIDAS Lab, IIIT-Delhi: YOLOv8 product-image extraction, 87% accuracy
B.Tech ECE @ IIIT Delhi
void reviewed this README · 0 bugs found · cards redrawn daily by github actions