Work

A healthcare research firm connecting patients with pharma and biotech companies for clinical and observational research had outgrown its no-code stack. 10,000+ patient participants across 55+ countries, no CI/CD, no version control on the frontend, no engineering practices. The platform needed to handle protected patient data, complex conditional logic, and longitudinal research participation.

Three applications, 459 of 592 commits in nine months. Auth systems, a compound-logic community builder, referral/rewards infrastructure, and configurable onboarding. Introduced CI/CD, branch strategy, and preview deployments from scratch.

Platform shifted from transactional patient engagements to continuous longitudinal participation. Operations team can configure workflows without engineering involvement.

A consulting agency's delivery model ran through the founder for every function — design, PM, engineering, client strategy. The no-code workflow created handoff friction, and the team needed engineering practices to scale delivery without scaling headcount.

Migrated development from no-code to code-first (Vite + Next.js) with CI/CD, branch strategy, and preview deployments. Introduced prototype branches that let stakeholders interact with features before committing to implementation — compressing the feedback loop from weeks to days and eliminating the design-to-implementation handoff. Built a multi-agent orchestration system for engineering workflows — intent classification, reusable prompt templates, subagent evaluation protocols, and agent contracts. Authored the structured planning methodology that drove all delivery: frozen specs → confirmed decisions → modular implementation streams → session logging. Built an MCP server for file-based prompt sequencing that inspired the shift to git-tracked, file-based knowledge infrastructure. Designed and built a headless agent runtime (~10.5K lines across 2 repos) with container-per-client isolation, scoped permissions, gated sub-agent spawning, context assembly from production APIs, and a training bench UI — the execution layer for the agency's AI consulting model.

Engineering methodology and orchestration system used for both client delivery and internal consulting work. The founder now prototypes and ships directly on git branches. Operations streamlined to a lean team focused on consulting and delivery.

A founder-led AI video generation startup was moving fast on a rapidly evolving codebase. Before a fundraising round, they needed clarity on what was actually in the system — implicit architectural decisions, structural risks, and where cognitive load was accumulating without documentation.

Structured audit of the V1 engine. Mapped architecture and data flow, surfaced undocumented decisions, identified fragility in the billing, prompting, and recovery semantic systems, and delivered a prioritized refactor and reorg proposal. Findings organized into a handoff document set for immediate use by the founding team.

Founder had architectural clarity before investor meetings. Clear map of what to stabilize vs. evolve — and a documented system the next engineer could actually onboard to.

Client had PDF mockups and spreadsheets. No technical spec, no designer, no team. Needed a production platform with thousands of unique content pieces that matched a specific author's voice and style.

Full-stack platform (Next.js, MongoDB, OpenAI, S3, serverless functions) with an NLG pipeline that reverse-engineered stylistic influences into structured prompts. Designed prompt chains with varying output lengths, input combinations, and formats. Built Node scripts that scrape, prompt, validate, and seed the database. Solo across ~253 hours, frontend through deployment.

Functional MVP with 18+ report sections, daily content generation pipeline, and early monetization built into Phase 1.

Musicians don't need another generation tool. They need something that listens to what they've actually recorded and helps them think about it — harmony, rhythm, texture, phrasing. Every 'AI music tool' on the market generates output from nothing. None of them start from the musician's own work.

A real-time voice-driven agent that analyzes uploaded tracks through a multi-stage pipeline: stem separation (Demucs), per-stem pitch tracking and transcription, phrase boundary detection, and rhythmic analysis. The user selects a section, speaks a request, and the agent generates musical options rendered as audio with notation. Built on Next.js, FastAPI, PostgreSQL, S3, SQS-driven worker chain with GPU inference endpoints, WebSocket agent communication, and Whisper/TTS for voice interaction.

Full pipeline operational: stem separation, per-stem analysis, real-time voice agent, and musical synthesis. Hardening and AWS deployment underway — v0 launching soon.