I'm a Mechanical Engineering student and product-minded software developer building software that makes complex workflows simpler and more reliable.

I primarily build interactive learning products and browser-first developer utilities, with a focus on client-side state architecture, input parsing, and automated verification. Alongside web engineering, I run empirical reinforcement learning experiments.

Portfolio · LinkedIn · Email (Hi@manav4u.tech) · X (@ManavShips)

A zero-account study tracker and spaced-review workspace for SPPU engineering students.

• Architecture & state: Browser-local persistence (localStorage) backed by schema-validated JSON backup/restore (v3.1.0 → v4.2.0 migration and one-click rollback) plus a pure spaced-repetition queue derived directly from syllabus progress and activity logs.
• Verification: 114 unit assertions across 14 Node native TypeScript test suites and 11 Playwright E2E / axe-core accessibility specs running in GitHub Actions.
• Links: Live App · Source Code · Architecture & ADRs · Latest Release

A 15-chapter visual Git and GitHub course with an in-browser practice terminal, interactive concept replicas, and a 243-command searchable cheat sheet.

• Architecture & state: Custom 22-command in-browser Git state machine (virtual working tree, staging index, branch pointers, and live commit DAG) verified against 61 real git CLI outputs; zero-backend cross-device progress portability using shareable #gv= snapshots and monotonic state merging; network-first offline PWA service worker.
• Links: Live Course · Engineering Case Study & Representative Test Vectors (application source is private)

A client-side developer workbench that parses, cleans, explains, and converts cURL commands into request code across 18 language and library targets.

• Architecture & analysis: Shell-aware lexer (Bash, PowerShell, and cmd.exe continuations, ANSI-C $'...' strings, glued short flags) lowering to a typed ParsedCurl intermediate representation; static analyzer that classifies 25+ browser-generated headers, preserves server-sensitive origin/referer headers for review, and runs heuristic credential detection before sharing.
• Links: Open Toolpaper · Engineering Case Study & Parser Vectors (application source is private)

An empirical study of reward hacking (specification gaming) vs. potential-difference reward shaping on Gymnasium's LunarLander-v3 using Stable-Baselines3 PPO.

• Experiment & evaluation: Compares an untrained baseline, standard PPO (269.41 ± 14.90), a flawed airborne-hover bonus (-56.07 ± 20.20 true task score), and a frame-to-frame potential shaping wrapper (ΔΦ = Φ(sₜ) − Φ(sₜ₋₁) − 0.05). All checkpoints are evaluated with training wrappers disabled, including a 500-episode deterministic benchmark (274.98 mean reward, 495/500 episodes ≥ 200).
• Links: Source, Checkpoints & Evaluation CLI

Merged

• Quantinuum/guppylang#2517 — Compile and lower comptime functions with fully qualified names, with integration test coverage across module boundaries.
• rapidfuzz/RapidFuzz#505 — Honor falsey callable processor arguments in cdist and cpdist across Cython and Python fallback paths, with 30 regression test cases.
• jquast/wcwidth#316 — Preserve terminal SGR styles when dropping whitespace-only wrapped lines (maintainer-integrated follow-up PR #316, which carried forward five commits authored in my original #310 submission, alongside maintainer follow-up changes).
• tree-sitter-python-orchard#15 — Give bitwise AND higher operator precedence than XOR in the Python Tree-sitter grammar (merged on Codeberg).

Currently Under Review

• lepture/mistune#552 — Python 3.15 regex scanner compatibility fix.

Issue Investigation

• colinhacks/zod#6743 — Reproduced and traced a false boolean-schema $ref resolution bug in JSON Schema conversion, providing a minimal reproduction and proposed regression test.

Building browser-first developer tools, deepening software engineering fundamentals, and contributing bug fixes to established open-source projects.

I optimize for understandable architecture, predictable state management, and reliable user-data handling. I use AI-assisted development selectively, validating implementation decisions through deterministic tests and edge-case analysis.