🌐 Portfolio
https://subhranshu.com

📎 Resume
Subhranshu's Resume 2026

🤝 Hire me
work.suhbranshu@gmail.com

▸ Hello! Myself Subhranshu.

▸ AI & Backend Engineer specializing in production LLM systems — multi-agent orchestration, RAG, semantic search, MLOps, LLMOps — and the high-throughput, distributed Python backend services & data pipelines that run them at scale.

▸ As the founding software engineer at an AI startup, I owned architecture and delivery across the stack, building POCs and systems that scale with both users and ambition, with stakeholder communication and user thinking. My work helped us raise a $4M seed round, and land partnerships with NVIDIA, Microsoft and Accenture.

▸ Roles I'm interested in — AI Engineer, GenAI Engineer, Backend Engineer, Fullstack Engineer, Product Engineer

▸ Currently pursuing a part-time MS in Computer Science at Georgia Tech (OMSCS).

▸ Technologies:

• GenAI: LangChain, LangGraph, LangSmith, RAG, Multi-Agent Systems, Semantic/Hybrid Search, Reranking, LLM Evaluation, OpenAI, Anthropic, Vertex AI, Azure AI Foundry
• Backend: Python, FastAPI, REST, SSE, WebSockets, Airflow, Celery, Microservices
• Databases: PostgreSQL, MongoDB, Redis, Neo4j, OpenSearch, Milvus, Pinecone, Supabase
• Cloud & DevOps: GCP, Azure, Docker, Kubernetes, Terraform, ArgoCD, GitHub Actions, Helm
• Frontend: TypeScript, React, Next.js, Redux Toolkit, React Query, TailwindCSS, Shadcn
• Observability & Testing: OpenTelemetry, Prometheus, Grafana, LangSmith, Sentry, Playwright, Pytest

▸ Resume: https://dub.subhranshu.com/Resume

▸ Portfolio: https://subhranshu.com/

▸ LinkedIn: LinkedIn/subhranshu-pati

▸ Email: work.suhbranshu@gmail.com

An agentic RAG platform for nuclear regulatory compliance — from the idea stage, through a $4M seed, into Prometheus, a $60M DOE Genesis Mission Phase II program with Idaho National Laboratory.

• Architected and shipped a production Multi-Agent RAG system with a knowledge base of 10M+ PDF files, using LangGraph & LangChain, that reduced nuclear compliance drafting from weeks to hours, using Claude Opus for the agent and Claude Sonnet/Haiku for sub-agents/tasks.
• Designed the Agent Layer: hallucination guardrails, routing, handoff, parallelism, human-in-the-loop, agent memory (both long and short-term), reflection loop and error recovery, token budgeting, prompt caching, resulting in generation of reliable and accurate citation-backed regulatory documents (1,000+ pages).
• Drove a Search Engine overhaul, implementing Hybrid Search (BM25 + Vector Search + RRF) using OpenSearch, that reduced retrieval from 3-4s to under 200ms, without any compromise in accuracy, by using Cohere Rerank, combining sparse & dense vectors and metadata pre-filtering.
• Owned the Document Ingestion Pipeline end-to-end using Airflow hosted on GCP, so customers could onboard terabytes of data without losing records or needing engineers to babysit failures.
• Delivered a secure, highly available, scalable REST Backend using Python FastAPI with SSE and Redis Caching and Streaming, that preserved user's progress, in case they disconnected in middle of agent's answer streaming.
• Established an MLOps pipeline to train & deploy a YOLO model for circuit diagram extraction, using Roboflow, Vertex AI, and Weights & Biases, auto-validating newer model against the older deployment, preventing regression.
• Introduced a LLM-as-judge Evaluation Framework for our Agent, using LangSmith, comparing any new changes to the Agent against a golden dataset, to evaluate — faithfulness, helpfulness, correctness and relevance, that ensured changes were non-regressive and followed TDD approach.
• Led the CI/CD buildout with GitHub Actions and ArgoCD on GKE, deploying MongoDB, OpenSearch, our backend, and Airflow on Docker & Kubernetes for scale, with Terraform for IaC, auto-deploying every code change with zero downtime and one-click rollback if a release fails.
• Engineered Custom Extraction and Chunking Systems, with image extraction, table structure parsing using NumPy, and contextual chunking, and using OpenAI's text-embedding-3-small, eliminating the information loss that drives most wrong answers downstream.
• Ensured reliability of production AI systems through end-to-end tracing in LangSmith, plus OpenTelemetry, Prometheus, and Grafana observability (LLMOps).

• Built a production grade Home Search Platform that processed voice described lifestyle preferences into structured search criteria, scaling relevance based ranking across 50,000+ property listings, using Pinecone Vector Database and LangChain for the Voice Chatbot.

• Embeddable citation-grounded RAG chatbot with per-tenant Milvus Vector DB isolation, event-driven Celery ingestion, and an async FastAPI backend with SSE streaming.
• Isolation is physical, not a WHERE tenant_id = ? — every tenant gets their own collection, making a cross-tenant leak structurally impossible instead of conditionally prevented.

• MCP server exposing a self-hosted Proxmox homelab to Claude: Home Assistant control, domain-scoped sub-servers, STDIO bridged to remote HTTP, using FastMCP.
• CasaOS + Portainer running open-source replacements for iCloud, Google Suite, S3 and Bitwarden, public surface via Cloudflare Tunnel, private access over Tailscale, and zero open inbound ports.

• A composable skills layer on top of Obsidian that holds projects, goals and habits in context and acts on them, built with Claude Code — not a to-do app with an LLM stapled on.

GenAI


Backend & Data



Cloud & DevOps



Frontend


Observability & Testing

• Video Games(Multiplayer FPS, RPG, 3D/2D Platform)

• Music(Old School Hip-hop, Dubstep, City Pop, Jazz, Future Bass)

• Video Editing and 3D Modelling(Blender,Davichi Resolve)

• Reading(Sci-Fi, Autobiography)

LinkTree
Portfolio
Twitter
LinkedIn