Building in public · 2026

AI products that do
the tedious part for you.

I'm Mohit, an Android & Embedded engineer turned AI product builder. I ship small, sharp, production-grade AI services: retrieval over messy official sources, one-tap summaries, and structured extraction from everyday email.

Selected work

03 projects
01 — RAG · Agents

ImmigrantAI

Official immigration rules, made human.
Live

Thinking about moving to Australia, Canada, Singapore, Luxembourg, Portugal or Dubai? ImmigrantAI turns official government immigration websites into clear step-by-step guides, checklists and fee tables, refreshed automatically every day.

Ask questions in plain language and get answers that cite the official source, or let the advisor suggest the best country, the visas that fit you, and a personal path to get there. Sign in and it remembers your details so every answer is tailored to you.

Highlights

  • Daily scrape → diff → classify → chunk → embed pipeline on Fly.io
  • Hybrid search: pgvector + full-text, fused with RRF and visa-scoped
  • Two-tier chat router: fast model for simple asks, deep model for hard ones
  • LLM judge + keep-or-restore gate guards extracted steps from regressions
  • Per-user memory and answer cache, all behind Postgres row-level security

Stack

Next.jsSupabasepgvectorPython workerFly.ioOpenRouterFirecrawlVercel
02 — Summarization · Distribution

summrize.it

Any YouTube video or article, summarized in one tap.
Live at dijest.my

Share a YouTube video or web page to "Summarize It" and get a point-by-point summary in seconds: a themed summary page, a rich WhatsApp preview card, and a plain-text version ready to paste into any chat or email.

Built to be cheap, fast and reliable: aggressive caching and dedup, a tiered multi-provider LLM fallback ladder, and a self-hosted transcript service so YouTube summaries don't break.

Highlights

  • iOS Shortcut + Android share flow, plus a dijest.my/<url> prefix
  • Server-rendered 1200×900 Open Graph cards for crisp chat previews
  • Provider ladder across Gemini, OpenRouter, OpenAI and Anthropic
  • Redis as the content store: summaries persist and rebuild instantly
  • First-party analytics groundwork for monetization

Stack

Python / FlaskVercel serverlessUpstash RedisPillowtrafilaturaFly.io
03 — Extraction · Mobile

AllMyShopping

Every online order, in one place.
In development

A mobile-first app that tracks your online orders across every shopping app in India and the US: what arrives today, what's being returned, and whether each refund has actually landed.

Orders flow in from forwarded emails, screenshots or the share sheet; an AI pipeline classifies each message, extracts structured order data, and merges follow-ups into the right order instead of creating duplicates.

Highlights

  • Deliveries by day, order timelines, and a returns & refunds tracker
  • "Refund overdue" and "return window closing" alerts
  • Per-user forwarding address via Cloudflare Email Workers
  • Vision-capable extraction, measured on a golden set (≥90% field accuracy target)
  • Installable PWA today, native apps via Capacitor next

Stack

Next.js 16Tailwind 4SupabaseCloudflarePlaywrightOpenRouter

How I build

principles
Cheap by design

Caching, dedup and right-sized models first. Every model and provider choice is tracked against cost.

Measured, not vibed

Golden sets, evals and A/B tests decide which model ships. Live end-to-end tests run against production.

Secure by default

Row-level security in the database, server-only keys, and guardrails that refuse to ship a regression.