Gerzain Oledo Nakamura
AI Solutions Architectfor Health & Learning.
I design, build and run AI systems where privacy is a product requirement, not an afterthought. Clinical AI, anonymous learning platforms, autonomous agents.
Open to senior roles and partnerships in Canada: AI architecture, health-tech and learning platforms. Based in Toronto.
- Zero learner identities stored, verified by audit
- 635,029 clinical records migrated with privacy built in
- 83% lower agent operating cost, from the cost records
Start here
Zero
personal data of learners persisted, verified by audit
An anonymous learning platform in production for Johns Hopkins
Nobody can know who took the course; the client still gets the completion analytics it has to report.
Try it
Two pieces that run in your browser on synthetic data. No server, no tracking.
The privacy threshold in action
Counts below N are hidden, rates disappear when either side of the fraction is hidden, and the export inherits the mask.
Concordance calculator
Agreement between an AI and experts, with its confidence interval and the human ceiling next to it.
How I work
- 01
Privacy in the code, not in the policy.
If a system must not know who you are, the data model, the headers, the logs and the rate limits are built so that it cannot. A privacy policy describes; the code enforces. The anonymous learning platform persists zero personal data of learners, verified by an audit of real responses and real database documents.
- 02
Human judgment decides.
In clinical work the AI suggests and pre-fills; the professional decides, and the AI never speaks to the person seeking help. An autonomous agent gets a human gate before it touches a credential. This is a design constraint written into the system, not a disclaimer.
- 03
Measure honestly, publish the ceiling.
A number without its interval and its human ceiling misleads in both directions. When an AI system is validated, the labeling is blind, every labeler stays in the result, and the report says when a target is met by the point estimate but not by the lower bound. Counts that could identify people below a threshold are hidden, and rates exist only when both sides of the fraction clear it.
- 04
Every decision on record, with what was turned down.
Each project keeps a decision log: the option taken, the alternatives rejected and why. Superseding a decision is done in writing. The anonymous platform went live with 25 decisions on record, one of them superseded cleanly; the client received the log.
- 05
The whole system, not the piece.
Architecture, data, compliance, infrastructure, cost, backups, a runbook and the client relationship, delivered together. When a project needs it, the content layer too: instructional design, video and audio, interactive screens. The client sees data and reports; the system keeps running without me in the room.
- 06
No lock-in. The client owns its content and its data.
Open standards where they exist (SCORM, xAPI, LTI), packages served untouched so the client can take its course elsewhere, models behind a router with budgets so a provider can be swapped, and code and data that transfer at the end of an engagement.
Where I have done it
Three domains, each with what was built. Figures come from the project's own records; the cases carry the method.
Health
- A clinical ecosystem for TERCET, a specialty clinic in Mexico City: patient module and CRM, 635,029 clinical records and 12,825 patients migrated, health-record and privacy compliance built into the product, passwordless access with anti-enumeration, 679 automated tests.
- A clinical AI system for a national mental health helpline in Mexico, run by Fundación Origen: data pipeline, risk detector with versioned prompts, and a blind validation study designed with the helpline's own counselors. Results are reported to the client first; the case is published once the client has them.
- Online programs for clinicians and families with UNICEF México, IMSS, the National Institute of Public Health and Cinvestav.
Learning
- An anonymous learning platform in production for Johns Hopkins, where nobody can know who took the course and the client still gets its completion analytics.
- An AI tutor with a live video avatar for Rise Real Estate Academy in Calgary, embedded in the academy's Moodle through LTI 1.3 and grounded in the licensing textbooks with retrieval-augmented generation. In production.
- A Moodle LMS operated at up to 45,000 active users, with programs for clinical, public-sector and corporate audiences; three MOOCs redesigned with AI for CLIMSS, the IMSS learning platform; a five-module course on maternal and infant nutrition in emergencies for UNICEF, the National Institute of Public Health and UAM.
Systems and data
- An autonomous operations agent with layered memory, scheduled jobs for the mechanical work and a human gate on credentials, migrated across three cloud providers and four model generations in five months, with an 83% reduction in operating cost measured from the cost records.
- A portfolio method: every project under one status board, a canonical registry of deadlines with alarms, numbered audit findings, and a compiled wiki of reusable patterns.
- An internal communication community for the HR area of Televisa, launched and run in production.
Who I have worked with
Clients and institutions, named without logos. Nobody on their teams is named anywhere on this site.
- Johns Hopkins Bloomberg School of Public Health, Moore Center
- UNICEF México
- IMSS, Fundación IMSS and CLIMSS
- Instituto Nacional de Salud Pública
- UAM Xochimilco
- Cinvestav
- Fundación Origen, Línea de Ayuda Origen
- TERCET
- Rise Real Estate Academy
- Grupo BAL
- Consejo de la Judicatura Federal
- Televisa
- TecMilenio
- BBVA
- Pineda Covalín
- UCB
Credentials
Google Cloud Certified: Generative AI Leader
Google Cloud
Verify on Credly · Certificate ID
229fa4b4bfba40549efde2494319b4c5Google Cloud Associate Cloud EngineerIn progress
Google Cloud
Data Analytics Certificate
BrainStation, Toronto
Diploma in Project Management (PMP-aligned)
Institute of Project Management
Diploma, Desarrollo de Aplicaciones para Negocios
Tecnológico de Monterrey
Bachelor's degree, Marketing
Universidad Tecnológica de México
Who I am
I like the problems that have no recipe: the ones that mix sensitive data, institutions that do not speak the same language, and technology that changes every month. I build the whole solution, document it so that it lives without me, and leave the last word to human judgment. Before software I produced audio and stories, and it shows: my courses are staged, not lectured. I work in English and Spanish.
Contact
No form, no tracking.