I build AI-poweredsoftware that ships.

Most AI coding tools are a single assistant bolted onto an editor. They autocomplete, but they don't own a change end to end: planning, building, reviewing, securing, and shipping it.

A full AI software-engineering platform built around a multi-agent pipeline: Planner → Repository Analyzer → Architect → Engineer → Reviewer → Security → Performance, each agent confirming its work before the next begins. Helix indexes and embeds a whole repo on connect, generates reviewable diffs in a real editor, and keeps a line-by-line intent ledger linking every generated line back to the request, plan, and tests. A self-verifying agent runs builds and tests in a sandbox and fixes errors before handing work back, then deploys one-click to an edge runtime with logs and rollback.

Live at helixstudio.org with an autoplaying product demo. Built on Next.js 16, React 19, Prisma 7, and Vercel Sandbox, with model-routed Anthropic + OpenAI agents and 24 built-in engineering skills.

Lecture recordings are where studying goes to die. A two-hour video with no structure means scrubbing around hoping to spot the whiteboard change, and transcription tools hand back a wall of text with no map.

One Next.js app carries the entire backend (36 API routes) with an Expo mobile app as a pure client. Uploads go straight from the browser to Vercel Blob, then a crash-tolerant pipeline extracts audio with ffmpeg and transcribes in parallel Whisper chunks with per-chunk database checkpoints, so when a serverless function dies at 300 seconds the job resumes instead of paying Whisper twice. Chapter timestamps come from matching verbatim quotes against the transcript, because gpt-4o drifts up to 100 seconds if you just ask it for times. YouTube imports skip transcription entirely by pulling the caption track, and when YouTube bot-checks datacenter IPs, the mobile app fetches captions on its own residential IP and hands them to the server.

Live in production on Vercel with Stripe billing: metered transcription minutes, idempotent charging so retries never double-bill, and four-hour lectures that chunk, checkpoint, and chapter themselves without babysitting.

Medicaid home-care fraud is simple in kind and massive in volume: visits billed from the wrong place, two caregivers sharing one phone, shifts that overlap in ways physics disagrees with. States pay first and chase fraud later, and a score that says 87 with no reasons is useless in a hearing.

A NestJS + Prisma backend where six pure-function detectors (impossible travel, shared device, duplicate visits, GPS anomaly, identity mismatch, abnormal duration) feed a weighted noisy-OR fusion, so each detector's share of the final 0 to 100 score falls out of the math and investigators see exactly what drove an 87. Every decision is hashed with SHA-256 and stored append-only, row-level security is forced on 19 tables so a forgotten where-clause returns zero rows instead of another tenant's data, and 150 Terraform-defined AWS resources (ECS Fargate, RDS, WORM evidence storage) are ready to plan-and-apply.

A foundation-stage build that is honest about it: 27 Prisma models, 34 partitioned tables, versioned detectors with passing tests, a 28-path OpenAPI contract, and an investigator dashboard. The hard product decisions (explainable scoring, humans in the loop for adverse action) are made and written down.

Travelers need quick utilities (currency, units, tips, time zones) exactly when they have no signal. Most apps break the moment you go offline.

A cross-platform Expo / Expo Router app that's offline-first by design: it ships with a bundled exchange-rate snapshot, silently refreshes and caches fresh rates on-device when online, and falls back to the latest known rates (with a "rates as of…" note) when offline. A growing grid of tools (currency and unit converters, tip and split, world clock, phrasebook, flashlight and SOS) driven by a single tool catalog, so new tools drop in fast.

Building toward a Play Store release with EAS, including in-app purchases via RevenueCat. One Expo app, shipped to iOS, Android, and web.

ION Water's client intake relied on manual, error-prone paper forms that were slow to process.

Led front-end development with React/Next.js and Tailwind, integrated with a SQL backend. Implemented dynamic routing, QR-code generation, and secure API handling with environment variables; built reusable components and mobile-responsive layouts.

Shipped on Vercel as a 4-person capstone team at the University of Louisville, working through security, CORS, and backend integration challenges.

Students juggle deadlines scattered across syllabi and PDFs with no single view of what's due.

Django web app that parses uploaded Excel/PDF schedules and auto-populates an interactive calendar via JavaScript. Added dynamic task management and an automated email-reminder system using Django background tasks (Celery/Crontab).

Deployed via GitHub and Vercel. An early end-to-end build spanning parsing, scheduling, and backend automation.