Cracked dev. nerding out.
|
I'm Rohit — a backend & AI engineer from Kolkata. I build backends that hold up — distributed services, realtime systems and LLM agents with guardrails, evals and fallbacks under the hood. I care about the parts nobody sees: service boundaries, queues, caches, rate limits, auth, and what happens when a provider goes down. I build AI into systems, not around them. Computer Science (IoT, Cybersecurity & Blockchain) at Heritage Institute of Technology · CGPA 9.02 · Smart India Hackathon 2025 national finalist. Things I keep circling back to:
|
|
| 711 | API tests on Walkthru's agent platform |
| 13,311 | deep-sky objects scored live per request in SkyGuide AI |
| 24 | tools on a remote MCP server I built |
| 20 Hz | sensor stream driving a realtime state machine |
| 23 / 23 | attack-probe checks passing on a Go server I hardened |
🧭 Walkthru — AI agent platform
A LangGraph browser agent that runs real user journeys inside your own Chrome. The API stays stateless (interrupt() / resume over HTTP, a Postgres checkpointer); jobs run from a durable Postgres queue (FOR UPDATE SKIP LOCKED, leases, jittered backoff); a circuit breaker sits across the model providers; the crawler is SSRF-hardened; and a remote MCP server exposes 24 tools. Every safety rule is enforced in code, never in the prompt. 711 API tests, plus LangSmith evals on adversarial fixture sites.
LangGraph · FastAPI · PostgreSQL · Supabase (RLS) · MCP · LangSmith · Chrome MV3
🗺️ TripVerse — agentic planner
Two LangGraph planning graphs write into one normalized trip document. Budget, routing cost and day capacity are deterministic code; models only propose. Graph updates stream over FastAPI, with per-node model routing and a Groq → Groq → Gemini fallback on rate limits. 188 backend tests. Repo
LangGraph · FastAPI · PostgreSQL · SQLAlchemy · asyncpg · Groq · Gemini
🔭 SkyGuide AI — distributed realtime system
Three services: React, an Express 5 gateway and a private FastAPI astronomy engine behind an allowlisted proxy. It scores 13,311 deep-sky objects in one vectorised pass for your exact telescope and sky, and guides the scope live over a 20 Hz WebSocket stream with a hysteresis lock state machine. Expensive responses come from a stale-while-revalidate cache with single-flight. Repo
Microservices · FastAPI · Express 5 · Socket.IO · Astropy · MongoDB
🧠 Neuron — multimodal retrieval & RAG
Text, image and audio fused into one CLIP query vector and re-ranked by a weighted blend of BM25, dense and visual similarity. Location-aware RAG cites the exact page and line, and the backend degrades gracefully: lazy models, then BM25, then extractive answers.
FastAPI · CLIP · Whisper · FAISS · BM25 · MongoDB
🔖 Shiori — Go server hardening
I security-audited a Go/Echo media server (~300 routes) and closed a critical unauthenticated network exposure. I added a per-IP rate limiter, lockout, constant-time token checks and a strict CSP, then built sandboxed JavaScript providers (Goja) with automatic failover on top. 23/23 attack probes pass.
Go · Echo · GORM · SQLite · Goja · WebSockets
💬 Yap Chat — realtime messaging & WebRTC
Room-scoped realtime chat and 8-way peer-to-peer video. Socket identity is verified from a JWT, and the WebRTC mesh is glare-free and never routes media through the server. 35/35 tests pass against a live server. Repo
Node.js · Socket.IO · WebRTC · MongoDB · JWT
A quiet, text-led portfolio in five themes. A die picks one on each visit. It holds case studies of everything above, my writing, and what I'm working on now.
Vite · GSAP · Lenis · Vanilla JS
| Area | What that looks like |
|---|---|
| Agent harnesses | The runtime around a model: step budgets, interrupts, tool gates and safe stops enforced in code |
| Sandboxes | Running agent actions and untrusted code with least privilege across browser, process and network boundaries |
| Agent memory | Checkpointed state, short- and long-term memory, what an agent should recall across sessions |
| Retrieval pipelines | Hybrid search, chunking that preserves citations, deciding what reaches the model and at what cost |
| Evals & observability | LangSmith traces, adversarial fixtures, regression suites: proof an agent works, not a demo |
| Distributed systems | Queues, idempotency, circuit breakers and caches that keep AI features up when providers aren't |
I took the long road to LLMs — NumPy, Pandas and scikit-learn, classical ML, deep learning, then transformers and modern GPT architectures — and I'm still working through the maths underneath: linear algebra, probability, optimisation. Models aren't black boxes if you know what's inside.
AI / LLM: LangGraph · LangChain · LangSmith · RAG · MCP · tool calling · embeddings · FAISS · evals Systems: microservices · API gateways · caching · rate limiting · job queues · circuit breakers · sharding & partitioning Security: JWT & sessions · OWASP Top 10 · SSRF / CSRF defence · row-level security · threat modelling
- Smart India Hackathon 2025 — National Finalist (backend for LivestockIQ)
- SBI Ideathon 2025 — Top 1000 nationally (technical lead, PraanaCare)
- Why does JavaScript feel so absurd? — Part 1: type coercion & IEEE 754
- Why does JavaScript feel so absurd? — Part 2: functions, scope & composition
- Why does JavaScript feel so absurd? — Part 3: closures, the TDZ &
this
The goal isn't just to build applications with AI.
It's to build the systems around AI — and make them hold up:
Models → Tools → Agents → Memory → Orchestration → Evals → Production
That's the rabbit hole I'm currently in.
心臓を捧げよ
Shinzou wo Sasageyo