Solo-built the production SaaS at qlabaqua.com — 38 REST API routes across 30 live pages in React/Next.js TypeScript. The interesting part is the money path: an idempotent Razorpay core where unique-index upserts and lost-race recovery make retries safe, timing-safe HMAC verification over raw webhook bodies, and Python jobs reconciling the ledger every 30 days.

Fail-closed Postgres rate limiting holds transactional state at 500+ concurrent requests, with an independent edge-CSRF guard. GitHub Actions CI gates on 330+ Vitest unit tests — including a 43-case payments suite — and Zod validates at every boundary.

→ the full system, dissected: FIG. A.1

Built therocketmedia.tech from scratch solo — a Next.js static site on Sanity CMS, 98 pages served via Cloudflare — and wrote the serverless waitlist API whose SHA-256 email digests dedupe repeat signups to one write. A custom GROQ cache-busting wrapper defeats Cloudflare Pages’ build cache, which had been shipping stale content.

Then the GenAI side: an 11-stage, 3,900-line Python pipeline in Docker — gating, backoff retries, checkpoint resume, renders to AWS S3 — producing 8–12 reels per episode. INT8-quantized Whisper large-v3 transcribes at around 3% WER, force-aligned with wav2vec2; LLM output is gated through 12 Pydantic schemas with fuzzy repair, and 30–90 s clips are clustered by cosine similarity (0.8 cutoff), merged via union-find with LLM tie-breaks.

The selection funnel is measured, not asserted: 11 of 75 candidates shipped as reels (≈15%), with rejection analytics logged to SQL.

Mapped machine-learning opportunities across Delhi Transport Corporation’s 44-depot network and 2,000-plus bus fleet, through structured interviews with 15-plus officials across six departments. Eight recognized problems became three shortlisted, solution-designed projects — transfer optimization, e-ticket sales growth, range prediction — presented to senior DTC leadership as a transformation plan targeting 50%-plus operational-efficiency gains.