Built / verified
BrandOss: persistent editorial workflow. Spokenly: a private, idempotent corpus. VisionarIAs: research turned into messages and infographics, with human review.
Founder Engineer / Applied AI
Jeaneth
Hernández
Ríos
I build AI products—and the layer that decides whether they can move forward.
Evidence, limits, cost, privacy, and human approval.
Software Engineer III @ Elsevier · Founder & CTO @ Mi Primera Chamba
Mexico City
Scroll to begin
01 / The editable internet · 2011
I wanted my Tumblr to look different.
I didn't learn JavaScript because I had a career plan. I wanted my Tumblr to do things it wasn't built to do.
So I opened the code. That's when I learned something that still drives my work: the internet wasn't finished.
02 / Designing the interface
First, I changed it for myself. Then, for other people.
I started by making the internet feel like mine. Later, I learned to shape it around someone else's needs.
A user.
A business.
A product.
03 / Engineering
The interface wasn't enough anymore.
I wanted to know what happened behind it: where the data went, how systems failed, how they scaled, and which controls made a release safe.
30+
critical vulnerabilities removed
↓80%
deployment time
2018→
from design to full-stack systems
04 / The AI shift
Models propose. Rules decide.
I build agentic systems where models speed up the work, gates limit what can move forward, and people keep the final say.
2022202420252026
05 / Recent systems · by evidence status
Building also means saying what's still missing.
A local test isn't production. An architecture isn't a pilot. I keep those lines visible.
Prototype / local validation
MXTW Planner organizes hundreds of events with organization-level privacy. MotoVolt explains compatibility and uncertainty before recommending physical validation.
Designed / specified
FlotaLista models a multimodal waste-collection network. There is no real pilot or validated local savings yet.
06 / Product · Mi Primera Chamba
Mi Primera Chamba
An AI job marketplace for people looking for their first role in Mexico. I built a chain of trust: ingestion, deduplication, explainable matching, faithful documents, and spending limits before a model call.
Ingest & verify
Multi-source ingestion, deduplication, three layers of fraud signals, and prompt-injection sanitization.
Explainable matching
A–F matching with visible gaps. No black-box scores.
Generate & deliver
CVs and letters have fidelity limits that block unsupported claims. Delivery includes a signed PDF.
Trust & economics
Spending is capped before a model call. Limits fail closed; prompt regression and golden-set evals protect fidelity.
Privacy by design
Candidate data is protected under Mexican data law (LFPDPPP).
In production
A real pipeline takes a raw vacancy to a delivered application, with traceability at every step.
STACK — Next.js 16, TypeScript, Supabase (Postgres, pgvector, RLS), Gemini, pg-boss, Python, Google Cloud Run, Vercel.
[COMPLETAR] Revalidate ~1,500 live vacancies and beta status before publishing those figures.
07 / Current role · Elsevier
Elsevier
Software Engineer III—supporting technical leadership in Elsevier's first AI-driven development pilot team.
I own the AI development harness and critical features. I also run workshops and support the team through the transition to AI development, while remaining accountable for architecture, accessibility, tests, and quality before merge.
AI development harnessCritical features
Workshops & enablementAI transition support
Full-stack deliveryAccessibility
Authentication & authorizationAutomated testing
Workflow governanceArchitecture review
08 / Lab · Positronica Labs
Positronica Labs
A lab where questions about agents, privacy, spending, and automation become small products people can actually test.
VenusAI
Conversational AI on WhatsApp. In production according to the supplied profile; current clients need revalidation.
Aldava
A CDMX rental tracker with a dashboard, API, discovery workers, and a Telegram bot. Local demo; deployment still pending.
ByteCoven
Monitors 100+ engineering blogs and sends digests through WhatsApp and Telegram.
katsi
A local-first MCP server: relational codebase context without sending the repository off the machine.
TallyGraph
In progress: visible AI spending and secret blocking before a cloud-model call.
Hypeman
In progress: an editorial desk with traceable variants and publishing that always requires human approval.
09 / Skills
I build. I verify. I bring it to people.
My most useful skill isn't a framework. It's turning uncertainty into a system people can test, explain, and operate.
I build
Mi Primera Chamba and VenusAI connect product, models, data, and operations.
Elsevier and MPC: search, ingestion, authentication, and traceable delivery.
katsi keeps code context local-first; PartProof combines agents with rules.
From product design to accessibility as a delivery requirement at Elsevier.
I verify
MPC runs prompt regression to block merges that break fidelity.
RLS, prompt injection, LFPDPPP, and 30+ vulnerabilities removed at Solera.
A GitHub Actions and Docker pipeline cut deployment time by 80%.
MPC caps spending before a model call.
I bring it to people
MotoVolt and FlotaLista make clear what works and what still needs validation.
Elsevier's AIDD pilot: harness, critical features, workshops, and support; MPC's technical process.
BrandOss and VisionarIAs turn evidence into reviewable work.
Morras en IA and hackathon mentoring bring the learning back to the community.
CORE: Python · TypeScript · React · Next.js · Node.js · PostgreSQL · Supabase · pgvector · RAG · MCP · LangGraph · multi-agent systems · evals · CI/CD
APPLIED: Redis · BullMQ · pg-boss · SQLite · Neo4j · Docker · Cloud Run · Vercel · GitHub Actions · Vitest · Jest · Tailwind · Three.js
PREVIOUS: C#/.NET · PHP/Laravel · Django · Robot Framework · Appium · Selenium · Figma · Adobe XD · Illustrator
LANGUAGES: Native Spanish · Professional English · Limited working French
Models propose. I build what decides whether they can move forward.
10 / PartProof · the trust layer
Public prototype / sandbox
I built PartProof's decision core. An agent can propose a purchase for a 3D-printing farm, but it can't skip evidence, exact compatibility, budget, or human approval.
VERIFIED LOCALLY — deterministic gate, typecheck, 84 tests—one skipped—and a public demo that returned HTTP 200 when checked on Oct 20, 2026.
[COMPLETAR] Verify the award before writing “we won.”
Seven-rule gate
- Evidence required
- Exact compatibility required
- Variant & shipping verified
- Price cap capped
- Budget available
- Quote validity active
- Human approval passkey
AI proposes. I build the infrastructure that decides what can move forward.
Building AI for LATAM? Let's talk.