I build and operate products end to end.

Forward-deployed at Policlinico El Bosque in Lima: I embedded in their operations, mapped the workflow and decision-making, then designed and rolled out a pre-arrival digital intake system, winning buy-in over two competing options. I trained the teams and ran a 30-day post-launch KPI cadence with leadership. Wait time down 50% (20 to 10 minutes), no-shows down ~40%, no added staff.

A multi-vendor integration was quietly leaking bad data. I found the breaks, built validation at every join, and put it all in one view leadership could actually read. Mismatches down 25%, manual reporting down 40%, breaks caught at ingest instead of weeks later.

Independent client work, ongoing. Building AI sales agents for Taos Global Group, a firm spanning insurance, accounting, and investments, to solve cross-selling across divisions and lead qualification under three different regulatory frameworks, plus the CRM and automated email campaigns behind it. Built to California's rules for automated sales calls: bot disclosure under SB 1001 and call-recording consent under CIPA, not just the federal baseline.

Built on 824K real transaction line items, this stress-tests marketplace unit economics: take rate, fulfillment cost, promotional funding split, and demand assumptions, then solves for the exact incremental volume a promotion needs to break even on contribution margin. A decision simulator, not a demand forecast: every assumption is a slider you can see and challenge, not a hidden model.

Real IRS filings for 20 nonprofits, turned into operational risk signals: reserve runway, leverage, and revenue concentration, screened against adjustable, explicitly-labeled thresholds rather than treated as universal standards.

Sentra started as v1, built as Product Manager and team lead in IBM's SkillsBuild AI Experiential Learning Lab, a four-month MVP delivery program where I led a 4-person team building an oversight layer for autonomous agents: policy enforcement, cumulative risk tracking, and a three-strike shutdown, demoed to IBM at program close. I adapted it into v2 for an NYU Berlin team project I led: medication-specific runtime governance, middleware between an AI prescription system and the pharmacy that validates every AI-generated order against live patient data and returns allow, block, or escalate in under 100 milliseconds, with an audit trail built to the EU AI Act. We presented v2 to Google and Join Capital and won the final pitch before a faculty panel.

A value-based-care problem taken to a live, working product. It ranks open care gaps by clinical urgency, response likelihood, and equity, then drafts personalized patient outreach with Claude (using prompt caching), behind a public methodology card. Applied AI where the prioritization has to be defensible, not a black box.

Two AI models check each other's work so you don't have to: a worker writes, an overseer stress-tests it, and agents respawn with memory after repeated failures. A practical take on AI reliability, catching the bad output before it ships.

A photo of the affected area plus a symptom description in, schema-valid pre-visit triage JSON out: ranked causes, red flags, and questions for the doctor, built on Claude's vision API. It ships with a side-by-side eval harness comparing image-plus-text against text-only, because in clinical AI you prove the approach is better, you don't assume it. (v1 fine-tuned an open model with LoRA; v2 is the multi-modal build.)

A live tool that builds a virtual influencer from real choices, hair, skin tone, body type, outfit, personality, and voice, then gives her a photorealistic portrait, a written voice, a spoken voice, and a lip-synced talking video, chaining three separate AI systems into one consistent persona.

An in-progress build of a live conversational avatar, speech transcribed as you talk, a generated response, and lip-synced audio back through a photo, all in one loop, architected against the same real-time pattern Tavus and HeyGen use, self-hosted instead of built on their API.