AI, computer vision and data-at-scale systems, from architecture to production.

A systematic value investing platform that values 60,000+ companies a day over a 5 TB+ database.

The problem

Methodically and systematically valuing every company in the market at once, with daily recomputation on price changes plus quarterly/annual reports, without the system buckling.

Stack decision

Re-architecting the stack from MySQL to PostgreSQL and from Pandas to Polars to make real-time computation viable; billion-row tables with a partitioning scheme to optimize queries.

The technical challenge

Designing a generic data module that generalizes to any asset type and its edge cases (different markets, reporting frequencies) over a 5 TB+ database.

Impact

Real-time daily valuation of 60,000+ companies on a dedicated server with 300+ GB of RAM and 10+ TB of disk. The backbone of Baboon Technologies’ investment fund.

A WhatsApp chatbot that lets any club member manage their bookings by chatting, powered by an LLM and tooling.

The problem

Give any club user a natural way to book courts, check when the next ones are free or do any booking task, without apps or forms.

Stack decision

An LLM + tooling (function calling) mechanism: the model interprets the natural-language conversation and calls tools that query and operate the club’s booking system, all over WhatsApp.

The technical challenge

Turning free-form requests ("when can I play tomorrow afternoon?") into reliable actions on the booking system — resolving availability, slots and confirmations inside the chat itself.

Gesture-based highlight detection and player heatmaps from each match’s footage.

The problem

Automatically extracting a match’s best moments and player analytics from already-recorded footage, optimizing cost and processing time.

The technical challenge

Telling the chosen highlight gesture (raised arms) apart from similar in-game moves (smashes, volleys): an algorithm over the joints (YOLO Pose) that catches true positives without adding false ones.

Impact

Processed with GPU workers on RunPod to optimize cost and time, with a strong focus on multithreading.

An edge device (NUC/Raspberry) in each club that streams, records and uploads video, orchestrated over VPN.

The problem

Efficiently live-streaming and recording each club’s video, uploading it to the platform and knowing what every NUC should do at any moment.

The technical challenge

Building the VPN architecture (OpenVPN → WireGuard) and a centralized RTMPS server with NGINX to remotely monitor camera and process connectivity in each club.

A multi-client agentic chat engine with RAG, external tools and generative BI, isolated per client.

The problem

Give each client an assistant that answers any question about their documents and data, connects to their tools and generates charts and dashboards, while keeping data isolated.

Stack decision

FastAPI + LangChain/LangGraph, RAG indexing with pgvector and Cohere Rerank, orchestration of several LLMs by quality/price (OpenRouter, Gemini, Deepseek) and tools via MCP/FastMCP. Fully traceable with Langfuse.

The technical challenge

During indexing, automatically clustering information, detecting discrepancies and allowing manual curation; and isolating data by project_id in a reusable multi-tenant engine.

Impact

A reusable engine in production for several clients (financial Q&A, margin and profitability monitoring with charts).

A platform for at-home therapists with a slot optimizer that minimizes travel and waiting time.

The problem

Cutting travel and waiting time for at-home therapists, bringing several roles (therapists, patients, centers, guardians, admin), clinical records, calendar and payments into a single platform.

Stack decision

Next.js/React + FastAPI/SQLAlchemy on PostgreSQL; Stripe (recurring payments and Connect), Google Calendar (two-way sync) and OpenRouteService for real distances/times.

The technical challenge

The slot optimizer: it scores (0–100) each therapist’s gaps and places every patient in the best one using real locations, penalizing dead waiting time and travel.

The cross-cutting infrastructure that orchestrates, deploys and monitors every service and project.

The problem

Managing multiple services and platforms in a reproducible, well-monitored way, with continuous deployment, backups and alerts.

Stack decision

Dokploy + Docker for orchestration, Traefik as reverse proxy, DNS on Cloudflare, WireGuard VPN, SSH tunnels with socat, daily S3 backups and job monitoring with Uptime Kuma plus centralized observability.

An MCP that navigates the 250+ server ecosystem to diagnose and fix remote incidents with the power of LLMs.

The problem

Diagnosing and fixing incidents across a 250+ server ecosystem where many resources are only reachable by hopping through the network: cameras accessible only from their remote server, and local databases with temporary data on each server.

Stack decision

An MCP (Model Context Protocol) that exposes tools to the LLM to navigate the network of servers and load balancers, inspect remote servers and cameras, query the local databases and apply fixes, on top of the video-analytics ecosystem’s 443 reverse-proxy/load-balancing architecture.

The technical challenge

Reaching deeply nested resources (cameras and local DBs only accessible by hopping server to server through the load balancers) and exposing that navigation as reliable tools so the LLM can diagnose and repair autonomously.

Impact

Monitoring and diagnosis of a 250+ server production ecosystem, speeding up the resolution of errors on remote servers and cameras.

The first crowd-counting system homologated by the Spanish Metrology Center, with >99% accuracy and real-time multi-camera fusion.

The problem

Certifiably accurate crowd counting in massive venues (trade fairs, casinos, street festivals), where a single camera can’t cover wide entrances and overlap between cameras causes double counting.

Stack decision

Python prototype (YOLOv2 + a custom tracking algorithm) rewritten in C++ with CUDA to scale performance; multithreading to process many cameras on a single GPU.

The technical challenge

Fusing and synchronizing up to 12 aligned cameras in real time into a single view where analytics run, with fine overlap configuration so the same person is never counted twice.

Impact

Over 99% accuracy homologated by the Spanish Metrology Center — the first certified counting system. Robust with up to 200 people under a single camera (San Fermín festival).

Turnstile counting, fraud detection and platform occupancy deployed across the entire Renfe network.

The problem

Monitoring accesses and platforms across an entire rail network: turnstile counting, fraud (two people at once, jumping the gate), people carrying objects, track intrusions and occupancy.

Impact

Deployed across the whole Renfe network and periodically audited by the client, recurrently validating its accuracy.

Final degree project: a 3D camera and a robotic arm that detects an object, approaches it and grabs it.

The problem

Assembling a 3D camera with a motorized robotic arm and a vision algorithm to detect an object and its distance, follow it, approach and grab it automatically.

The technical challenge

The 3D camera loses depth below 0.5 m; solved by interpolating with wheel odometry to estimate position in the near range, plus the arm’s inverse kinematics.

An electromechanical dispenser that logs and authorizes morphine extraction via fingerprint.

The problem

In an emergency, staff took morphine products without leaving a record; regulators penalized this because everything had to be logged.

The technical challenge

Coordinating a full electromechanical system: an Arduino board with sensors across 40 slots, a stepper motor, a servo-driven hatch and a fingerprint reader, all tied to a tablet (Java app + MySQL).