Case study 02 · 0-to-1 AI platform
After onboarding 29 distributors by hand, I saw the same bottlenecks on every account. I vibe-coded a working prototype on my own initiative to earn the mandate, then built the real platform from scratch with one engineer - writing production code in Claude Code.
Before the platform, onboarding ran across Google Sheets, Monday.com, Jira, Slack, and email with no single system. Business-logic capture happened through ad-hoc calls and Google Forms. Data mapping was manual. Jira tickets were created by hand. Onboarding took 152–191 days depending on account size, and the company needed to scale from 30 onboardings/quarter to 70 parallel onboardings to reach 300 distributors within a year.
The bottlenecks were specific:
In August 2025, when AI developer tools were just taking off, I vibe-coded a working prototype of an onboarding tool on my own initiative - not to build the real thing, but to show leadership how fast building had become and to pitch the idea. It worked. The prototype earned me the mandate to lead product for a new onboarding platform, moving me from hands-on onboarding into product leadership.
Every pattern, bottleneck, and workaround I'd seen across 34 manual onboardings shaped the platform's design. I built it from scratch with one engineer (Gayanga Kuruppu), using Claude Code for velocity.
A purpose-built questionnaire system with branching logic by ERP type, screenshots and visual examples on UI-specific questions, and "why we need this" context on every question to reduce confusion. I separated pay from business-logic intake after feedback that pay work continues well after ordering begins.
Once a distributor completes questionnaires, the system generates a structured specification as the shared source of truth, replacing manual spec writing. Delta reviews; the distributor signs off.
Instead of manually creating 30–80 tickets per account, a three-phase AI pipeline reads the signed-off spec and generates draft tickets by category for review and approval before creation.
An integration progress tracker showing ticket completion by category, a launch timeline flowchart, and progress bars - killing the dead zone where accounts stalled silently.
AI-generated "Story So Far" narratives, game plans, and launch-status classifications replaced ~25 min/account/week of manual synthesis across 43 accounts, triggered automatically from meeting approvals, at 98% grounding accuracy.
Shipped 74 PRs and 44K lines of production code in 9 months, written with Claude Code alongside one engineer.
| Metric | Before | After |
|---|---|---|
| Time-to-revenue | 189 days | 113 days (~40% faster) |
| Requirements / data capture (at launch) | 100 days | 26 days (~4x faster) |
| Intake median (Q2 2026, 32 accounts live) | 100 days | 8.5 days (11x faster) |
| Onboarding throughput vs. new bookings | ~0.4x (Q3 2025) | 2.0x (Q1 2026), ~$250K backlog cleared |
An analysis of 2,520 Jira comments across 703 tickets in 15 epics showed 23% of all comments were clarification requests. The account with the best spec quality also had the lowest clarification rate (11% vs. 23% average) - direct evidence that better intake reduces integration delays.