Cracked dev. nerding out.
Agent Peeps: a tiny live sticker for every Claude Code, Codex, Copilot and Antigravity session, so you can see what each AI coding agent is doing, and which one needs you.
npx agentpeeps initLive demo · Repo · npm · ⭐ if it's useful
|
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:
|
|
| 16 kHz / 24 kHz | PCM16 streaming pipeline with interruption flushing in KOE |
| 10 min | persisted approval expiry with atomic claims on agentic writes |
| 188 | backend tests on TripVerse's deterministic planner |
| 35 / 35 | live-server WebSocket & WebRTC tests passing in Yap Chat |
| 5 | standalone design systems driven by one central content schema |
| 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 |
| 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 & Agents: LangGraph · LangChain · LangSmith · Gemini Live · MCP · RAG · AudioWorklet / Web Audio · Tool Calling · Embeddings · FAISS · Evals
- Backend & Realtime: FastAPI · Node.js · Express 5 · WebSockets · Socket.IO · WebRTC · REST APIs · Pydantic · Asyncio
- Distributed Systems: Microservices · API Gateways · Stale-While-Revalidate Caching · Rate Limiting · Job Queues (
SKIP LOCKED) · Circuit Breakers · Sharding & Partitioning - Data & Storage: PostgreSQL · MongoDB Atlas · MySQL · Redis · SQLite · Supabase (RLS) · GORM · SQLAlchemy / asyncpg
- Security & Reliability: JWT & HttpOnly Sessions · OWASP Top 10 · SSRF / CSRF Defense · Constant-time Crypto · Sandboxed Runtimes (Goja) · Threat Modeling
- Testing & Tooling: Docker · Linux · Jenkins CI · Git / GitHub · pytest · Vitest · LangSmith Regression Suites
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





