Aadarsh Upadhyay

hold to see the picture

software engineer and architect — react, typescript, css, electron, node

I build React products end to end, and review my own pull requests with a bot I wrote. AI assistants move me fast; the review loop is what keeps the quality.

hi@capad.fyi see what shipped

now searchts 0.14.0 · GlyphMaps 1.0.0 · Grove v0.2.0-alpha

India · remote-first · open to relocate

My assistants don't get the last word.

A confident wrong answer is worse than an honest error, and that is how I build. On ComplyV the AI chat never mutates data: it proposes, you confirm, and a separate authenticated endpoint re-checks auth, company access and the business rules before anything is written — with an idempotency key and a line in a SHA-256 append-only audit log. Due dates and penalties come only from the deadline engine, never from a model.

On my own repos, Hare reads the diff and the CI as evidence and posts findings as comments. 127 app reviews on searchts, and not one approval: the event is hardcoded to COMMENT.

propose, then confirm

The model writes a proposal card. A person confirms. The write goes through an endpoint that checks permissions again, idempotently.

code decides, model suggests

The draft linter is code, not model self-grading. Eval cases bait number leakage on conceptual answers.

evidence, never instructions

PR text, diffs, commits and CI are evidence. Anything in them asking for an approval, a secret or a push gets quoted as an attack, not obeyed.

fail loud

searchts refuses to pretend. If it hits a login wall or an interactive CAPTCHA it says so instead of returning a shell that looks like a page.

languagesTypeScript · JavaScript · Python

productReact · Next.js · React Router · Vite · Electron · Tauri · Android SDK

front & designCSS · design systems · UI/UX · Figma · Tailwind

serverNode.js · Express · MongoDB / Mongoose · REST · JWT / RBAC · WebSocket

systemsmulti-tenancy · authz · append-only audit logs · agent tool design · eval harnesses

shipsPyPI · GitHub Actions · Cloudflare Workers / OpenNext · Docker · Tailscale

agentsMCP · Grok / Grok Bots · Claude · Claude Code skills · AGENTS.md · RUNBOOK

What shipped.

Every line is a thing that exists on the internet with a link you can open. Where a project is alpha or contributed, it says so.

searchts
shipped · MIT

A keyless web unlocker for AI agents: reads, searches and transcribes what a naive fetch cannot, from bot-walled pages to whole AI-chat share conversations. Ships on PyPI through 0.14.0 as a CLI, an MCP server and a Python library.
- 0API keys
- 3unlock tiers
- 8AI-chat providers
- code
- pypi
Hare
ongoing · not a SaaS yet

A PR reviewer in two forms. The bot checks the branch out, installs it, runs the repo's own test command and posts findings as me. The GitHub App runs the same contract on my own pull requests. Neither can approve a merge.

grounded in CI and diff evidence, never blocking
- 127app reviews
- 25bot reviews
- 0blocking
- pausing itself
- all app reviews
GlyphMaps
shipped · AGPL

Google Maps turn-by-turn on the back of the phone: the next maneuver on the Nothing Phone (4a) Pro's 137-LED Glyph Matrix, so a glance at a face-down phone shows your next turn. No Maps API key, no network code.
- 137LEDs
- 12maneuvers
- 2.3 MBAPK
- code
- download apk
Grove
alpha · GPL-3.0

A free Git review companion that sits beside your AI coding editor. Commit graph, diffs and every in-flight worktree, refreshing live as your agent changes the repo under you.

now: markdown fidelity + the large-commit freeze (v0.2.0-alpha)
- alphastage
- 1worktree per task
- BYOagent
- code
- windows alpha

Smaller, and not the headline. Still mine, still linked.

alpha

Dooper — a self-hostable universal chat client: every messaging network in one native inbox, instant sync, nothing behind a paywall. Tauri 2 and React over a Synapse homeserver and the mautrix bridges. code

contributed

wmux — not my project. A terminal multiplexer froze whenever I opened a diff pane on a non-git folder; I profiled the main process, found 92 percent of self-time in synchronous fs calls under a 2-second poll, and took one call from timeout to 0.7 ms. Four merged PRs: #135, #138, #153, #258. PR #135 all my PRs

ongoing

burncard — know what your AI coding agents actually cost. Reads your Claude Code and Codex logs on your own machine, so the numbers are yours and nothing gets proxied through anyone else. github npx burncard

also

MSU Halls Register — an alumni-register design prototype for MSU Baroda's 16 halls; live, engagement paused. msu.capad.fyi
CoffeeBreath — a Rainmeter music widget that breathes with the song's album art.
Discord Voice Overlay — a glass desktop overlay for live voice control.

The record.

Roles, dates and the systems behind them. Commit share is real: I was the primary engineer on ComplyV, not the only one.

Software Engineer and Architect

ComplyV, formerly Compliance Sarathi · Appson Technologies

Jan 2026 — 31 Jul 2026

Corporate-governance SaaS: MCA/ROC company data, statutory deadlines and penalties, resolution and minutes drafting, and Vee, an agentic chat with governed writes.

~553 of 647 commits, about 84 percent. Small team; not sole developer.
Propose-then-confirm writes, three-tier truth wall, multi-tenancy as an ownership-and-grant graph over CIN rather than a naive tenantId.
~1,400 LOC deadline engine for multiple MCA forms; 34 Mongoose models; ~40 suites / ~2,852 tests.
Fixed an LLP fee-slab bug and a DIR-3 KYC triennial rule that would have been wrong from 2027.

Junior Software Engineer

Wordibly · Appson Technologies

Jun 2025 — early 2026

US transcription, translation and captioning platform across human, hybrid and AI tiers.

Owned tops-transcript-editor, a Vite/React editor with find/replace and related editing flows, at roughly 107–113 commits.
Sprint delivery with no visual regressions on production-bound UI.

AI Training Engineer

Turing, via Appson Technologies

2025 · about two weeks

Producing Meta Llama training data and evaluation sets.

educationBachelor of Computer Applications (Honours), The Maharaja Sayajirao University of Baroda — expected 2028

fellowshipProduct Space PM Fellowship, Top Fellow, 2025–2026

What people wrote.

Verbatim, with the link each one was written in. The ellipsis in the first one is the only cut I made.

"#133, which is one of the best bug reports this project has had. ... You attached a debugger to the main process, took a CPU profile, and came back with 92% of self-time in readdirSync/statSync/readFileSync under walkDir/readCurrentFile, plus a before/after table showing system.identify going from timeout to 0.7 ms."

amirlehmam · Maintainer, wmux · source
"from your own IP" is the whole insight. a proxy pool fights the bot-wall; your own IP is already through it, same reason a human's browser doesn't trip cloudflare. an agent acting from where you already are doesn't need to sneak in. nice build.

Phi Browser (@phibrowser) · on searchts · source
fetch-time + final_url turns read output into something a reviewer can cite later. Tier/status is useful; redirect + timestamp makes it durable.

Dang_nh (@hikariraina) · on searchts · source
Aadarsh owned the architecture of Compliance Sarathi end to end and shipped a reliable agentic assistant under real deadline pressure. He is who you want on the hard parts of a system.

Engineering, Compliance Sarathi · Appson Technologies

Contact

hi@capad.fyi

Tell me the system and the deadline. India, remote-first, open to relocate; outside-India roles with visa sponsorship welcome.