← Aman Kumar Limited · RegisterNotes to the accounts · UNBR
Flagship segment · product · 2026 · Live
Unbordr
Schengen visa readiness for Indian applicants. AI checks every document, builds the checklist, fills the form.
- Role
- Building it — product, AI pipeline, backend, frontend
- Lead
- Outcome
- Live at unbordr.in — taken from idea to launch in 7 months, with 97% of the commits written by me.
- Links
- Live ↗
- Stack
- Next.js 16 · React 19 · FastAPI · Supabase Postgres · Claude (Haiku + Opus) · Razorpay · Redis · Railway · Vercel · Capacitor
Problem
Indian applicants assemble Schengen files by hand. Refusals come from a missing page, an old statement or a name spelled two ways. Templates don't help, because the rules depend on who you are: employed or self-employed, sponsored or not, first trip or fifth.
What I built
Unbordr runs on conditional intelligence: your answers decide which document rules apply.
- Two-tier document AI. Claude Haiku extracts every upload. Opus re-reads only when a deterministic check fails: a passport MRZ, an Aadhaar Verhoeff checksum, or low confidence.
- A separate verifier. An Opus pass works through a fixed list of checks, then looks across the whole case. Its verdicts feed back into per-check precision.
- A rule engine for checklists. Modules for rules, conditions, branches, substitutions and jurisdiction, not one big prompt.
- Form automation. The Schengen form is filled from case data. First I built HTML + Jinja2 tokens → Chromium print-to-PDF. Later I replaced it with overlays on the official PDF.
- "AI plans, code decides" itineraries. Opus plans each day from curated places only. A validator rejects any change to locked facts and falls back to a deterministic plan.
- Payments that unlock on the webhook. Razorpay orders, then staff-raised payment links. Paid features unlock only after the webhook confirms.
- An airline-themed product. A readiness "boarding pass", a flight route over a globe, and a paper-plane path along a great-circle arc.
Result
7 monthsIdea to live productsrc: git log + unbordr.in 97%Of all commits, written by mesrc: git log: 1,813 of 1,865 commitsUnder the hood
What I'd change
I'd have started with the advisor-assisted flow the product uses now, instead of a fully self-serve wizard first. Applicants wanted a human at the risky step.
Risk factors
- The product's shape is still moving: in October 2026 an advisor-assisted flow replaced the self-serve wizard.
- Extraction quality depends on upstream models; messy uploads escalate to the more expensive Opus pass.
This is a portfolio, not an offer of securities. No shares exist. Figures are about shipped work.