Seven years of shipping — from apps past a million downloads to the AI agent platforms I build today.
Six production systems built at AI Xccelerate, plus the independent products that got me there. Press play on any of them.
Animations illustrate how each system works — the real ones are under NDA
Multi-tenant voice and text agents on LiveKit: the full speech-to-text → model → speech pipeline, replies grounded in a private vector store, tool calls through one typed contract.
Voice AI Infrastructure · 2026
Replaced a per-minute vendor stack with a self-hosted realtime pipeline.
Typed tool contract shared by API and voice worker
A team wiki where one set of operations backs the human editor, the in-app agent sidebar, and an MCP server any external agent drives with a scoped API key.
Knowledge Infrastructure · 2026
One set of operations backs three clients: editor, agent sidebar, MCP server.
~360 tests, no network in CI
AI Xccelerate · 4 systems
One service turning agent output into branded documents, decks, single-page sites and microsites. An intermediate-representation architecture keeps the model out of the final markup, so what ships is deterministic.
Consolidated four in-house implementations into one IR-first artifact service.
A drop-in widget captures screenshots and recordings with technical context, has a model describe what went wrong, then routes it to per-project backlogs and back to the reporter.
Single embeddable file turns a screenshot into a triaged, routed bug report.
Email and LinkedIn sequences over contact enrichment, agent orchestration and deliverability monitoring, with organisation-level tenancy and row-level security enforced at the database, not the application.
Two outbound channels, email and LinkedIn, sequenced on one enrichment and deliverability stack.
A central catalogue of agent skills in an open, portable format: people author them in a web app; agents pull definitions over REST or MCP, scoped by tenant API key.
Agents search by task and pull full skill definitions over REST or MCP.
Self-shipped · 2020—2024
Web Platform · 2023
One AI-assisted pipeline generates and maintains the whole calculator catalogue and its sitemaps.
200+ tools
Android Application · 2022
Content, notifications and monetisation change from an admin panel, no app update.
50K+ downloads
Commercial Script · 2024
Sold as a self-hostable PHP script to other developers and small businesses.
Commercial licence
Mobile Application · 2024
A gradient interface system and admin-driven session content behind authenticated profiles.
Portfolio Site · 2024
Paintings first, and fast on the low-end phones the audience actually used.
Web Application · 2023
Realtime words-per-minute and accuracy scoring on a sub-frame keystroke input loop.
Archived means it shipped and ran for years — the domain has since lapsed, so it is listed rather than linked.
The parts that matter are the ones I can’t publish.
Happy to walk through any of these properly — architecture, trade-offs, and what broke first.
Karan Jadoun is an AI platform engineer based in Jaipur, India. He builds agent infrastructure — realtime voice platforms, retrieval systems and multi-tenant services — and has independently shipped mobile apps past 1.5 million downloads.
I started by shipping apps to people who had no reason to be patient with me. That is still the standard.
No CS degree, no team — a BCA, a laptop in Jaipur, and eighteen apps that crossed 1.5M downloads. Every bug was mine to find.
Voice pipelines, retrieval and MCP servers — agents as first-class clients, not bolt-ons.
Multi-tenant services with tenancy enforced in the database itself.
Design systems, motion that means something, and real accessibility.
Seven years taking ideas to shipped products, alone.
How I build
All three still running today
AI Platform Engineer
Developer & Digital Builder
Creator & Editor
Open to platform and agent-infrastructure work, and to the occasional product build worth doing properly.
Based in
Jaipur, India
GitHub
/karanjadounDetailed history and stack.