At Synthrun, I architected and shipped an autonomous multi-agent lead-ops system with an orchestrator, task queue, Firestore state, and Telegram human-in-the-loop workflows for live client automation. I also built a production mail system and a React webmail client.
At IIT (ISM) Dhanbad's HiSLM Project, I kept a two-node NVIDIA Jetson system running 24/7 using FastAPI and WebSocket request routing, load partitioning, systemd services, and Tailscale VPN. I cut QLoRA training loss by 37% and benchmarked quantised Qwen2.5 3B across hardware configurations.
I built Context Fence, a local firewall for AI agent tool calls that blocks secret leaks, flags hidden instructions, and applies YAML policies with SQLite audit logging. It makes decisions in under 10ms, and the project reached 71% trial-to-paid conversion with five paying customers.
At Outlier AI, I evaluated production LLM outputs and designed adversarial prompts to expose model failure modes for RLHF pipelines. I also designed and trained LSEM, a semantic similarity model using BiLSTM and multi-head attention, with Docker deployment and embedding-space visualisation.
