GenAI Backend Engineer
CS (AI & ML) student who builds GenAI backend systems and ships them — agent orchestration, RAG pipelines, reliable infrastructure.
Hi, I'm Subhamoy — I build production-style AI systems: RAG pipelines, LLM-powered backend architectures, and agent orchestration. I've published two MIT-licensed npm packages — rexa-agent (a Docker-sandboxed CLI coding agent) and @subhamoy/somoy (a provider-agnostic TypeScript agent SDK). I'm currently studying Computer Science (AI & ML) at Brainware University, and looking for an internship where this is the day-to-day work.
Here are a few things about me:
Building provider-agnostic model execution engines with Zod-typed tool I/O and typed RunResult failure states instead of exceptions. Implementing loop detection, agent handoffs with complete transcript transfer, and dynamic tool-calling loops with strict host vs sandbox git trust boundaries.
Enforcing isolated execution within non-root Docker sandboxes with strict command containment. Pausing execution for explicit user confirmation on consequential actions like git operations, dependency installs, and file deletions. Implementing input/output guardrails with automated secret scanning and OS-level credential storage via keytar.
Designing production LLM-powered backends: multi-modal ingestion pipelines across PDF, URL, YouTube, and raw text, pgvector for semantic similarity search, and asynchronous Redis/BullMQ embedding pipelines. Backing it with Redis-based rate limiting, high-throughput result caching, and Express/Node APIs built to hold up under real load.
Shipping and maintaining two MIT-licensed npm packages — rexa-agent and @subhamoy/somoy — from architecture through publishing. Deploying and running services on AWS, Vercel, and Render, with Docker Compose for containerized environments across development and production.
Autonomous CLI Coding Agent
Published npm package (rexa-agent) [ADD: npm weekly downloads]. TypeScript, Bun, Docker Compose, Gemini API, Tavily API, keytar. Runs in a non-root Docker sandbox [ADD: median sandbox startup time]; input/output guardrails with secret scanning; OS-level credential vault; pauses for confirmation on consequential actions (git ops, installs, deletes).
TypeScript AI Agent SDK
Published npm package (@subhamoy/somoy) [ADD: npm weekly downloads]. Provider-agnostic (Gemini, OpenAI, offline MockProvider) with Zod-typed tool I/O [ADD: per tool-call round-trip time], typed RunResult failure states instead of exceptions, loop detection, agent handoffs with transcript transfer.
RAG Document Intelligence Platform
PDF/URL/YouTube/text ingestion, pgvector similarity search [ADD: p95 query latency], async Redis/BullMQ embedding pipeline [ADD: documents ingested per minute], AWS S3, Clerk auth.
AI Resume Analysis Platform
Gemini-generated feedback, async BullMQ/Redis pipeline [ADD: average analysis turnaround time], AWS S3, Clerk auth.
Frontend Odyssey: The Interactive Web Experience Challenge
Indian Institute of Technology (IIT), Patna (hosted via Unstop)