B.Tech + M.Tech Dual Degree, CSE @ IIT Kharagpur · Kharagpur, India
I build systems close to the metal — matching engines, storage engines, bytecode VMs, network servers — and apply that same low-latency, reliability-first toolkit to AI and agent infrastructure: the layer that decides whether agentic systems are fast, correct and cheap in production.
- 🤖 Agent systems — multi-agent orchestration (LangGraph), hybrid RAG, sandboxed tool-use, evals & tracing
- ⚙️ Low-latency C++ — lock-free data structures,
io_uring,epoll, custom allocators, kernel-bypass I/O - 🗄️ Databases from scratch — B+ Tree indexes, ARIES WAL recovery, MVCC snapshot isolation
- 🏆 Competitive programming — Codeforces Expert, ICPC Regionalist
- Codeforces Expert, peak rating 1776 — global rank 118 in Round 1105 (Div. 2)
- ICPC Regionals 2026 qualifier
- IICPC QuantFest 2026 — Regionals Rank 347 (round co-titled by Optiver & Jane Street)
- Smart India Hackathon 2024 Finalist — Btrfs/XFS forensic tool recovering deleted inodes at ~200 MB/s
The hard part of agentic AI in production isn't the prompt — it's the runtime: grounded retrieval, isolated tool execution, self-correction, cost control, and observability. That's what I build.
| Project | What it is |
|---|---|
| my_ai_issue_resolver | Autonomous multi-agent pipeline (LangGraph FSM) that turns a GitHub issue into a reviewed PR — context → developer → tester → reviewer agents with confidence-gated re-querying, a Docker test sandbox, a self-correcting retry loop, per-issue token budgeting, Qdrant hybrid RAG (dense + BM25, AST chunking), plus LangSmith/Prometheus tracing and an evals harness |
| agentloop | A from-scratch, dependency-free agent runtime — a @tool decorator that auto-generates JSON schemas from type hints, a plan→act→observe loop with retries and token budgeting, pluggable providers (Anthropic + an offline mock), and fully inspectable trajectories. 35 offline tests, ruff-clean, CI on Python 3.10–3.12 |
| evalloop | An evaluation harness for agents (companion to agentloop) — task suites, deterministic checks (output / tool-use / efficiency / errors), an optional LLM-as-judge, and scored reports: success rate, tool-call precision/recall/F1, steps, tokens and latency. 36 offline tests, ruff-clean, CI matrix |
🔭 Currently building: an invisible macOS menu-bar AI assistant driven by the Claude agent bridge, and bringing agent-style tool-use into my systems projects.
| Project | Highlights |
|---|---|
| femto | Real-time order-matching engine (C++20, io_uring, LMAX Disruptor) — lock-free intrusive book, 780 ns p99 add-order latency, p99.99 tail cut 24 µs → 4.6 µs, differential-tested over 1 B replayed events in an ASan/UBSan CI matrix |
| htn · htn-server | Observable high-throughput service (C11, epoll, SO_REUSEPORT) — 42K RPS across 1,200 live clients, lock-free SPSC ring at 1.8M ops/sec, Prometheus + Grafana (lab under Prof. Sandip Chakraborty) |
| mini-sql-engine | Transactional storage engine (C++17) — B+ Tree over 8 KB pages with LRU-K buffer pool (point queries in 3 I/Os on 10 M rows), Volcano pipelining, ARIES WAL + MVCC snapshot isolation |
| ember | Language runtime (C++20) — hand-rolled lexer, Pratt parser, bytecode VM with computed-goto dispatch (180M ops/sec), incremental tri-color mark-sweep GC (p99 pause 1.8 ms) |
| Stat_Arb | C++17 pairs-trading backtester — lock-free ingestion, Kalman-filter hedge ratio, volatility-regime detection |
| EcoRoute | High-throughput C++ routing engine optimizing graph traversal for carbon emissions and AQI exposure |
Languages Python · C++ (11/17/20) · C · Java · SQL · Bash
AI / ML PyTorch · scikit-learn · NumPy · pandas · Matplotlib · LangGraph · RAG
Systems Linux · POSIX threads · epoll · io_uring · lock-free programming · low-latency · storage internals
Data & Observability PostgreSQL · SQLite · Redis · Qdrant · Prometheus · Grafana · perf
Testing & DevOps pytest · GoogleTest · Google Benchmark · ASan/UBSan/TSan · Docker · Git · CMake · GitHub Actions
Kharagpur, India