Software engineer at Honeywell (.NET / C# / React), working end to end across systems programming, backend services and applied AI.
I build the hard version on purpose — a search engine rather than a search wrapper, a VPN rather than a VPN client — and then measure it, because an unmeasured claim is just an opinion.
- 🔭 Now: Software Engineer @ Honeywell — .NET (C#), React, Kubernetes
- 🌱 Into: distributed systems, developer tooling, information retrieval
- 🧩 825+ LeetCode solved (130+ Hard)
- 🎓 B.Tech CSE, VIT-AP — CGPA 9.03/10
- 📫 msayantan05@gmail.com
Each of these is measured, not described. The numbers below are reproduced by a command in the repo.
| project | what it is | measured result |
|---|---|---|
argus Java 17 · zero deps |
A full-text search engine and a browser rendering engine, from scratch — inverted index, BM25, query DSL, WAL, plus HTML/CSS/layout/paint and a JavaScript interpreter | 98.9% of scoring work eliminated by Block-Max WAND with provably identical rankings · 218 tests · durability proven by killing a real JVM with halt(9) |
MyVPN Rust · QUIC · Wintun |
Serverless peer-to-peer VPN with NAT hole-punching, full-tunnel routing, kill-switch and DNS-leak hardening | Security controls unit-tested, not asserted — injection guard rejects 8 payloads, constant-time proof compare, domain-separated key derivation |
chromaforge C++17 · multithreaded |
Colour-grading engine — 3D LUT trilinear, separable convolution, ACES/Reinhard tone mapping, custom thread pool | 13.4× separable vs naive 4K blur · white balance 306 → 473 Mpix/s via lookup tables, bit-identical output |
GameManager NestJS · Socket.IO · Expo |
Server-authoritative multiplayer framework validated across 45 distinct rulesets, with per-player hidden-state projection and WebRTC voice | Crossplay web + Android, per-commit APK from CI |
VoiceRep Python · PyTorch |
Consent-aware, watermarked zero-shot voice cloning with an independent verification pipeline | Held-out LibriSpeech benchmark with acceptance gates independent of the engine's own score |
BitDown Rust · Tauri · axum |
BitTorrent client with DHT/PEX/UPnP and a byte-range HTTP server that streams a file while it downloads | Shipped Windows installer, auto-released from CI |
Three things these repos taught me that are worth more than the features:
- A randomised differential test found a bug no example test could. Argus scored the same document as
6.278620742691458on one path and...457on another — both scorers summed over aPriorityQueue, whose iteration order is unspecified, and floating-point addition is not associative. Rankings were non-deterministic across runs. Fixed by scoring in fixed clause order. - Work skipped ≠ time saved. Block-Max WAND eliminates 94.5% of scoring at top-10 but only runs 1.6× faster, because the postings are already in RAM so nothing is saved on I/O. Reporting both numbers is the honest version.
- A single timing is not a benchmark. On a hybrid P-core/E-core laptop, identical chromaforge runs disagreed by 30–50%. The harness now reports medians across five bursts and prints the spread.
Fixes sent upstream to libraries I use:
- bytesize (Rust) — fixed
f64precision loss when parsing large byte counts - go-humanize (Go) — exact integer parsing in
ParseBytes - luxon (JS) — reject out-of-range 12-hour values in
fromFormat
Open to SDE / new-grad roles · 2026