Five systems. Four countries. Shipped.

A U.S. veterinary chiropractic clinic network was running on a fragile WordPress + MySQL stack holding 2M+ medical records, with no mobile support, no offline mode, and no path to HIPAA compliance. Doctors were losing real time every day to record-keeping friction, and the system couldn't leave the browser.

• Single Flutter codebase shipping to Web, iOS, and Android — cutting development effort by roughly two-thirds versus three separate stacks.
• Firebase + Node.js + GCP backend (Pub/Sub, Cloud SQL) with offline-first sync for clinics in low-connectivity rural areas.
• Planned and executed an incremental migration of 2M+ medical records from WordPress / MySQL to the new platform, holding 99.99% uptime across a four-month rollout.
• HIPAA-compliant data handling end-to-end: access controls, audit logs, encryption at rest and in transit, security audits.
• CI/CD pipelines across dev / staging / production, shipping releases ~25% faster. Server query optimization cut response times ~30%.
• UI/UX overhaul that lifted retention by 15%.

Working beta in production within three months for the annual conference. Record-keeping time dropped ~3 minutes per entry — roughly one hour per doctor per day, ~300 clinician-hours saved collectively each week. Still the primary system in use.

Teams were drowning in WhatsApp threads, email, and Telegram — inbound messages lost in noise, outbound responses slow and inconsistent. They needed a single AI-powered platform that could handle intelligent messaging across channels, grounded in their own data, with real operational controls.

• Leading the project end-to-end — database design (PostgreSQL on Neon via Prisma), PR governance, and Cloudflare deployments across the team.
• Self-hosted Coolify infrastructure running microservices: WhatsApp integration, Telegram bot, and email responder, all plugging into the Wondercall AI core.
• Frontend on Cloudflare Pages; backend APIs on Cloudflare Workers with the Hono framework — edge-first performance, minimal cold starts, global reach.
• RAG pipeline with OCR ingestion, hybrid dense + keyword retrieval, reranking, and grounded refusals. Citation-linked answers so users can verify every response.
• Multi-tenant isolation for data, retrieval index, and model-call budgets.
• Eval fixtures and retrieval-quality regression tracking, so degradation surfaces before users notice.

Platform in production serving real inbound and outbound messaging workflows. Retrieval quality holds up on messy scanned-PDF corpora, and every AI answer links back to the source span.

Turbos, injectors, and fuel pumps are the parts that kill a trucker's month — expensive, failure-prone, and uncovered once the factory warranty lapses. DPS Protection sells coverage for them as a subscription, which makes the software the whole business: recurring billing, policy issuance, and a claims lifecycle nothing off-the-shelf models — replacement shipped out, failed core shipped back, deposit refunded on verification. Then the business model changed mid-build, from annual prepay to monthly coverage, and the platform had to change with it without breaking a single money path.

• The monthly coverage model: a recurring coverage fee per truck plus a fixed claim charge per replaced component, capped per 12-month term. Contract terms are frozen onto the policy at purchase — editing a plan never reprices a live agreement — with a 15-day grace period, 30-day termination, and admin reactivation that restores the original term so lapsing can never be used to unlock fresh coverage.
• Full claims lifecycle with money-safe gates: submitted → approved → claim charge and core deposit collected → shipped with tracking → core returned → deposit refunded. Every charge, refund, penalty, and credit is a ledger row; the shipment-release gate reconciles the invoice against the ledger before anything leaves the warehouse.
• Fleet checkout: N trucks in one application, each with its own plan and covered parts — one signature, one payment, one policy per truck, all-or-nothing with compensating refunds on partial card failure.
• Admin-managed catalog: plans, per-plan claimable parts, core deposits, coupons, and a fee-credit switch, all database-driven. Price changes mint immutable Stripe Prices; existing subscriptions migrate at renewal. Referenced products archive, never hard-delete.
• A 20-page admin panel (claims, customers, policies, payments, analytics, coupons, reviews, homepage/terms/privacy editing, editable email templates, team invites), a support-ticket workspace, and customer notifications over email and Twilio SMS.
• Serverless on Netlify: React SPA plus the entire Express 5 API as a single function, Neon Postgres with Drizzle migrations applied at deploy time, strict CSP, rate limiting — with the Express-on-Lambda edge cases (body parsing, real client-IP restoration) handled properly. Customer portal in English, Spanish, and Chinese.

Live at dpsprotection.com on real Stripe billing. The model migration shipped behind a test suite of 2,301 assertions with zero failures and a full end-to-end QA pass against real data on a dev environment before promotion. Pricing, coupons, claimable parts, notification templates, and site content are all self-serve — the client runs day-to-day operations without a developer in the loop.

A growing UAE e-commerce operation needed a production-grade platform with proper AWS infrastructure, a data model that could handle simple products, variants, and digital goods (license keys), and search that actually worked at catalog scale — plus back-office tooling so non-developers could manage the catalog.

• AWS infrastructure from scratch: CloudFormation templates, EC2 / ECS clusters, Lambda functions, Application Load Balancers, Route 53 DNS, and GitHub Actions CI/CD pipelines.
• Flexible product data model supporting simple products, variable products with variants, and digital goods — along with the relational structures to keep them consistent.
• Elasticsearch for performant full-text product search, tuned for relevance and autocomplete.
• SSR for SEO and first-paint speed; multi-language internationalization baked in from day one.
• Advanced filtering, ranking, and back-office APIs — including bulk CSV import / export so website managers can update catalog data at scale without touching engineering.

Production e-commerce platform running on AWS with CI/CD, full-text search, SSR, and a self-serve back office. Non-technical staff manage the catalog daily without engineering involvement.

A Korean fintech had a legacy Web3 wallet Flutter app that had ballooned to 80MB with no coding standards, no CI, and patchwork architecture. The team was shipping slowly and users were dropping off at install. Leadership needed the codebase restructured and the team's velocity restored.

• Overhauled the legacy codebase and cut app size from 80MB to 11MB — ~86% reduction — by auditing dependencies, removing dead code, and restructuring asset pipelines.
• Established coding standards and architectural patterns (MVVM + DI) that were adopted company-wide.
• Introduced TDD and CI/CD automation — tests + multi-environment deploy pipelines on what had been a manual, ad-hoc release process.
• As Project Leader, drove product planning and system architecture. Translated feature descriptions into technical specs and managed the engineering team.

Installable, maintainable Web3 wallet app at a fraction of the original size. Team velocity recovered, release cadence stabilized, and the architecture I introduced became the company standard.

Merchants wanted to accept Web3 payments without forcing customers into a specific wallet and without holding volatile assets. Off-the-shelf processors either didn't support Solana settlement paths cleanly or taxed the margin to death. A checkout SDK that handled the real-world edges — routing, staleness, failover — didn't exist.

• Wallet-agnostic checkout flow: customer pays in whichever SPL token they hold; merchant receives a chosen stablecoin.
• Jupiter integration for in-flight swap routing with slippage guards and quote-refresh windows.
• RPC provider abstraction with failover, cooldown, and jittered re-test — the service survives provider outages instead of dying with them.
• Real-time pricing with caching, staleness checks, and explicit reject paths when quotes move outside tolerance.
• Rate limiting, idempotency keys, and a settlement ledger that survives process restarts. Retries become reads, not re-runs.

Production Web3 checkout SDK handling live merchant payments. Per-transaction cost an order of magnitude below traditional processors; settlement finality in seconds, not days.