SYSTEM_ONLINE · Darbhanga, IN · UTC+5:30
I'm Mukesh Kumar Thakur — building AI systems for healthcare revenue cycle at RapidClaims. I design architecture, wrangle data pipelines, and make interfaces feel lightweight — recently cutting a front-end from 2 GB → 400 MB.
I care about the seam between infrastructure and interface — where performance meets polish. Over the last two years I've designed AI enrichment pipelines, rebuilt memory-leaking dashboards, and shaped the architecture for whole product modules from first commit to production.
Outside of shipping at RapidClaims I'm building an MCP marketplace — a natural-language control layer over Gmail, Calendar, Tasks, and Slack.
MCP marketplace + chatbot that speaks natural language to Gmail, Google Calendar, Tasks, and Slack. Multiple MCP servers, secure auth, automated workflows.
Scale-agnostic sync that ingests every message from any Gmail mailbox into Richpanel as a ticket — backfills decade-old inboxes and streams new mail in real time, fully deduped and threaded. Mailbox size is no longer a ceiling.
Streams every customer–agent conversation, classifies intent & sentiment, writes back to the product surface in near-real-time.
End-to-end returns orchestration with automatic label generation, workflow engine, and full operator tooling — shipped from first commit.
Hunted leaks across long-lived dashboards — from event listeners to zombie observers — and cut memory by ~80%.
A visual element picker for any webpage — hover, click, and the element lands in your clipboard as clean Markdown. A full HTML parser runs locally: headings, inline styles, links, images, fenced code with language detection, tables, nested lists, blockquotes. No servers, no API keys, zero data collected.
A password manager that never leaves the machine. Shortcut autofill, save-on-login, and Touch ID unlock via the WebAuthn PRF extension — the Secure Enclave returns a secret that wraps the vault's AES key, so the wrapped key on disk is useless without the authenticator. No account, no sync, no servers, no dependencies.
Conditional rendering as attributes — r-if, r-else-if and r-else sit on the element itself instead of nesting ternaries inside JSX. A Vite plugin rewrites them to standard React at build time, so nothing ships at runtime, with an ESLint plugin that flags broken chains in the editor and TypeScript types for autocomplete.
A tech-signal reader with a bar for entry. Every night it sweeps ~50 hand-picked engineering sources, has an LLM score every item against a curated interests profile, and serves only what clears the bar — each item with a two-sentence tldr. Multi-user accounts, topic filters, and votes that teach the ranking.
Raw files, readable at last — Markdown, JSON from any API response, and CSV open as clean documents instead of walls of text. A collapsible JSON tree that keeps 64-bit ids exact, search that finds matches inside collapsed nodes and shows each record whole, sortable CSV tables, twelve themes. Hand-written parsers, no dependencies, not a single network request.
Selected among students nationwide for Amazon's one-month ML program.
Built a game-style app about the seven Nimbus satellites in a 24-hr sprint.
Daily-DSA habit — data structures, graphs, DP, system-style rounds.
5-star rating across both tracks.
I'm open to senior IC roles, freelance collabs, and any problem where AI meets product. Reply usually under 24 hours.