Johann
Piedras

Building a multi-tenant inventory and point-of-sale platform for vintage resellers: Next.js 16 / TypeScript PWA, 81 API routes, a 22-model Postgres schema, role-based teams with scoped invitations and write-path audit logging.

Shipped Stripe Connect Accounts v2 direct charges with AES-256-GCM per-tenant credential encryption and idempotent processing across three signed webhook endpoints, plus a Python/FastAPI computer-vision sidecar (OpenCLIP ViT-B-32 embeddings, perceptual hashing, pgvector cosine search) split across Vercel and Railway, with graceful degradation so sales continue if ML is unavailable.

488 passing Vitest tests across 50 files, Sentry error tracking, PostHog session replay.

Built with Claude / Claude Code as a development accelerator. The architecture, data model, product decisions, and production tradeoffs are mine.

Founded and scaled a profitable multichannel vintage apparel business across a custom storefront, Depop, eBay, Whatnot, and regional markets, generating $60K+ in sales across 1,500+ items from a ~2,000-item inventory. Channel-level reporting on revenue, average selling price, sell-through, and inventory aging guides sourcing and pricing.

Scaled a Whatnot live-selling channel from ~$30K to $250K+ in monthly gross sales in under 12 months: show structure, curation, and standardized bulk listing workflows that cut listing prep ~60% and saved ~20 labor hours a week. Built show-level KPIs used across a ~30-person operation; trained and supervised ~5.

Led a commercial AI check-fraud platform from prototype through pilot to production handoff. OCR, computer vision, and NLP (AWS Textract, Azure Form Recognizer, TrOCR, EasyOCR, YOLOv7, Siamese networks) combined into explainable indicators for human reviewers. Pilots showed ~30% lower review time and a ~25% relative reduction in false-positive investigations. Led a team of five junior data scientists and contributors.

Analyzed 2M+ transaction rows with Python, SQL, and Tableau; built fraud-ring dashboards and recurring executive reporting presented directly to the VP and President of Fraud Analytics. Cut manual data prep ~40% while handling sensitive PII under GLBA/PCI controls.

Improved an internal Apache Solr search platform serving 10,000+ employees. Search-log, click-through, and zero-result analysis (TF-IDF, topic modeling, geo-targeting) contributed to a ~34–35% increase in platform usage.