I build complete systems.
From idea to production.

I'm Miguel León, a software engineer from Hidalgo, Mexico. This isn't a project list: it's a journey. Scroll slowly — everything I mention, you'll see.

Start the journey ↓ Write me
SCROLL ↓
#1
national — CENEVAL Award, best EGEL CS score
0
systems in production, today
0
businesses open their doors on my software
24/7
agents & pipelines running on their own
CHAPTER 01 — SEÑA

The dataset that doesn't exist — we're recording it

Mexican Sign Language has almost no data to train models on. Seña starts there: our own capture of hand, face and body keypoints + synced video, with real signers from the Deaf community and a portable capture studio.

Seña Studio — interfaz de etiquetado
MediaPipePythonPWA

Every video becomes a skeleton

From every recording we extract hands, face and body as keypoints: the original video, the overlay to verify, and the pure skeleton — the representation a model can learn from. The bridge between a real sign and trainable data.

Vistas original, overlay y esqueleto
21 keypoints/hand478 on faceVideo sync

And a way to learn it that hooks you

A lesson map with streaks and lives (yes, like Duolingo — but for LSM), an alphabet with a rotatable 3D hand, and practice with spaced repetition (FSRS, the algorithm behind Anki). Pedagogy validated with Deaf teachers. And every practice session feeds the dataset — learning is also donating data. That's the flywheel.

Mapa de lecciones de Seña Lección de abecedario con mano 3D
FSRSThree.jsLSM pedagogy

And it all points to the translator

Seña's core is a bidirectional LSM ⇄ Spanish translator, built in phases, each usable on its own: fingerspelling — a 29-letter classifier already training on our own data (67% accuracy and climbing, 8 signers) —, words and phrases from the seed vocabulary, sign → text with beam search + an LLM rewriting for fluency, and text → sign with real signer video, not avatars.

Métricas del modelo de Seña
29 letters + neutralBeam search + LLMData flywheel
seña · captura
Mapa de lecciones de Seña Lección de abecedario con mano 3D
CHAPTER 02 — SOMOS EMPLEABLES

An entire company runs on my code

I'm the lead engineer at Somos Empleables. SE Programa is the SaaS where the method lives: stages, journal, sessions and a custom CMS where the course is edited in blocks. Next.js 15 + Prisma + PostgreSQL, real payments via Webpay Plus (production, not sandbox), feature flags and three environments — with a full built-in CRM to run the client portfolio.

Hub del alumno en SE Programa
Next.js 15PrismaWebpay Plus
VISIT WEBSITE ↗

A CV studio with AI inside

The student sets a target role and the CV Studio generates a tailored résumé: an LLM writes using vector search over the client's history, outputs real LaTeX and compiles to PDF (or opens in Overleaf), tunable by pasting the job posting. AI applied to a real pain, not a decorative chatbot.

Estudio de CV con IA
LLMVector searchLaTeX → PDF

Headhunter: jobs that find you

Every student sees fresh offers for their target role — 322 today, updated this morning. Behind it: 26,464 vectorized job posts in Qdrant, 500+/day ingested from LinkedIn, and a CV ⇄ job similarity engine deciding what reaches whom.

Ofertas de empleo en SE
Qdrant · 26k vectorsEmbeddings500+ offers/day

I also built the infrastructure — and I run it

An ARM64 VPS I stood up from scratch and keep in production: 7 containers live, 4 SSL domains, 24 cron jobs, real CI/CD (push to main → production; branches → their own dev environment with a separate DB) and observability with Telegram alerts at $0 cost. The flow you see is the real scraping in self-hosted n8n: 40+ nodes feeding the Headhunter daily.

Flujo de n8n del scraping
7 containersCI/CD → prod24 cronsn8n

A full CRM — operated by agents

The admin is a real CRM: a Telegram message inbox, portfolio health traffic lights, bidirectional Notion mirror and 5 metric dashboards — 291 historical clients, 75% success rate. On top, a network of 6 specialized AI agents (staff, analytics, CRM, SRE, triage and processes) reporting on their own every week.

Dashboard admin de SE
CRM + Notion6 AI agents75% success
somosempleables.com
CHAPTER 03 — PRINCIPIA

I turned my life into a data system

Performance OS: a daily ETL merges my biometrics (WHOOP), activity (Strava), habits (Notion), screen time and chess into 365+ days of history — with an annual recovery heatmap, served via MCP to my AI agents, which answer questions about my own performance.

Vista Rendimiento de Performance OS
FastAPIMCPDaily ETL
SEE IT LIVE ↗

And the data tells you things you can't see

I don't just collect data: I interrogate it. The cleanest correlation in the dataset: less screen time, better-recovered nervous system. And exercise turned out to be the anchor habit — on training days I complete almost 4x more habits. The domino that knocks down the rest.

Correlaciones de hábitos
Screen → HRVAnchor habitn=356 days

11,601 hours of screen time, tamed

1,376 days of history, 518 apps tracked. From averaging 11 hours a day in 2023-2024 to half that today. Measuring isn't curiosity: measuring changes behavior.

Historial de pantalla
1,376 days518 apps−50% screen

Even my life fits in a dashboard

The Life view: every dot is a week, from primary school to today — contests, chess, IPN, the pandemic. And at night, a ritual with "Jung": an agent that listens to how the day went and gives perspective back. Sovereign data, even the sentimental kind.

Vida en semanas
Life in weeks"Jung" agentBrain dump

And it all runs on my own cloud

None of this lives in someone else's cloud: local ARM64 servers on my desk, with Docker and secure egress via Cloudflare Tunnel. Sovereign data, fixed cost, full control. From the same lab: a RAG agent with Nietzsche's complete corpus in Qdrant.

Servidor local
Docker ARM64Cloudflare TunnelQdrant
performance os · dashboard
CHAPTER 04 — CLIENTS

Software running real businesses

Clinical Space: the private-practice platform for Dr. León (anesthesiologist & intensivist) — scheduling, records and the clinic's daily operation, end-to-end.

Clinical Space
Medical platformMariaDBDocker
VER CÓDIGO ↗

From ticket to cash-out

Klub (events & tickets) and Tusados Shop (barbershop) — systems used every single day, maintenance included, delivered in weeks, not months.

Klub
E-commercePOSBookings
SEE CODE ↗

Orders that walk themselves to the kitchen

Garden 27: the point of sale for a café/bar — categorized menu, tables, orders straight to the kitchen, cash-outs and an admin panel. Built for tablet, used every shift.

POS de Garden 27
POSTables + kitchenCash-outs

I also teach — and I build my own tools

I teach chess to kids at Club Ajedrez Tlaxcoapan, and I built the class app too: 6 game modes (play, capture-me, pawn race, piece setup, puzzles and openings), student registry and today's class — running on a local server in the classroom, designed for iPad.

App del Club de Ajedrez
JavaScriptSQLiteDockeriPad

From code to the physical world

I come from physical engineering: a robotic arm in SolidWorks, the Tlaxcoapan mills modeled and 3D-printed, and the mobile workbench I built in high school — welding included. That's why I can take software into the real world: cameras, sensors, my own servers.

Brazo robótico CAD Molinos impresos en 3D Mesa de trabajo móvil
SolidWorks3D printingWelding
clinical space · dr. león
CHAPTER 05 — TRACK RECORD

The road to here.

// PROFESSIONAL EXPERIENCE
OCT 2025 — PRESENT

Somos Empleables

AI & Full-Stack Engineer — Technical Architect
  • se-programa: employability SaaS in production charging real payments (certified Webpay Plus) — unified stack for the low-ticket course and high-ticket coaching hub.
  • Full infrastructure on an ARM64 VPS: Docker Compose, Nginx, multi-domain SSL — hosting the platform, n8n, internal dashboard and public site.
  • Interview transcription pipeline (Whisper + n8n) for automated coaching feedback, with bidirectional Postgres ⇄ Notion sync.
  • Job scraping with Apify + n8n (40+ nodes, 500+ offers/day) feeding semantic matching on Qdrant.
Next.js 15PrismaPostgreSQLAuth.js v5Webpay Plusn8nWhisperQdrantDockerARM64
VISITAR WEB ↗
FEB 2025 — SEP 2025

BizBat

Data Science Intern
  • Normalization and unification of user and event data in BigQuery, with automated continuous-ingestion pipelines.
  • KPIs and reporting: led GA4 reporting and built interactive Looker Studio dashboards for strategic decisions.
  • GCP infrastructure: optimization and data preparation for downstream ML model training.
BigQueryGoogle CloudGA4Looker StudioPythonSQL
NOV 2024 — ABR 2025

Novolabs — Startup Incubator

Automation & Data Engineering
  • AI architecture: Whisper transcription processed by an LLM to score reading comprehension and generate feedback.
  • Automation of the gamified data flow with n8n for user analysis and reading improvement.
  • Business validation: led the model validation phase, market strategy and user research.
WhisperLLMn8nGamificationValidation
CHAPTER 06 — DIPLOMAS

The trophy case.

10 physical awards — click any of them for a closer look.

Premio CENEVAL
CENEVAL Award — #1 national EGEL CS score · 2024
1er lugar nacional Estadística
1st place national — Probability & Statistics
Concurso nacional de Cálculo
Top-10 national — Differential Calculus
Olimpiada Mexicana de Matemáticas
Mexican Math Olympiad — 1st place, state round · 2017
Concurso Pierre Fermat
National finalist — Pierre Fermat Math Contest
Excelencia académica IPN
1st place national "A la Cachi Cachi Porra" — IPN campuses
Campeón Sub-12 de ajedrez
U-12 champion — state qualifier, National Chess Olympiad
Campeonato Nacional de Ajedrez
National Open Chess Championship — representing Hidalgo
ONMAPS
National Mathematics Olympiad (ONMAPS)
Certificación lectura rápida
Certification — learning techniques & speed reading
Miguel León

Shall we build something together?

Web platforms, business systems and applied AI — or your grant proposal. Tell me what you need and I'll tell you how I'd build it, straight up.

Send me an email
SAME-DAY RESPONSE · HIDALGO, MX
Diploma ampliado