Projects

What I have built, with the code and, where there is one, a demo you can try. Each project explains the problem it solves and why I built it that way.

Own projects

Konjo

In progress

Konjo is for people who want to get more done and stay organised in a day full of notifications. Productivity apps tend to add more lists, alerts and settings when what is missing is focus. Konjo helps you decide what to do next, start without friction and keep going without relying on willpower alone. Coming soon to Google Play.

Learn more
  • Works offline: data lives in a SQLite database on the phone and the app never waits on the network. It syncs only the changes since the last sync with a Spring Boot backend (Java 21, PostgreSQL and Redis); if two changes conflict, the most recent one wins.
  • XP is stored as an append-only log of events, so the total always adds up and can be audited. On top of it sit 70 tiered achievements, and sleep, steps and heart rate are correlated with productivity.
  • The architecture and the decisions are mine; the agents work inside a seven-stage pipeline. Before any code is written, each change goes through design and a security review checked against the OWASP Top 10. Afterwards, a weighted score decides whether it moves on or goes back. The backend build fails below 80% line coverage, and a hook stops the agents from accessing sensitive data.
  • My own RAG served over MCP: a FAISS index of the 33 spec documents that rebuilds itself when one changes, so the agents look up the passage they need instead of loading a whole spec. In its benchmark it answered 15 of 15 queries with 72% fewer tokens.
  • Java 21
  • Spring Boot 3.5
  • PostgreSQL 17
  • Redis
  • Flyway
  • React Native
  • Expo
  • TypeScript
  • SQLite
  • RAG
  • MCP

SentiLife

Completed

When an elderly person falls, hours can pass before anyone finds out. SentiLife detects falls in real time with an event-driven architecture. I led the backend, built from written specs: REST APIs with Spring Security and JWT, Virtual Threads, RabbitMQ, and metrics in Prometheus and Grafana. Deployed on AWS with CI/CD.

  • Java 21
  • Spring Boot
  • RabbitMQ
  • PostgreSQL
  • Docker
  • AWS
  • Prometheus
  • Grafana
  • SDD

Legacy Modernization

In progress

A case study in understanding and modernising an undocumented legacy application, as if nobody who wrote it were still at the company. The base is an online shop in Java 11 with Spring MVC, MyBatis and JSP, which I work through with AI help: first as a user, then layer by layer down to the database. The user walkthrough and the architecture, with diagrams, are done. Next comes extracting the business rules, pinning them down with characterization tests and moving the catalogue to a new API using the strangler fig pattern.

  • Java 11
  • Spring MVC
  • MyBatis
  • JSP
  • Mermaid
  • Claude Code
  • Strangler fig

Hackathons and company challenges

Pledge

Top 10 of 250+ teams · HackMeridian 2026

A rental deposit sits idle in the landlord account for months, and getting it back depends on their goodwill. Pledge puts each deposit in its own escrow contract on Stellar: neither side can move it alone, it earns yield while the lease runs, and it goes back to the tenant once the landlord approves. Tenants never touch crypto: they sign in with a passkey (fingerprint or Face ID), pay no network fees and fund the deposit from their bank. Four of us built it in 28 hours; it runs on the Stellar testnet.

Learn more
  • If there is a disagreement, only the landlord can stop the refund, and only before approving it; Pledge staff then decide the split. If the landlord does nothing within 30 days of the lease ending, the tenant gets the whole deposit back.
  • The tenant signs with a passkey and a relayer submits the transaction and pays the fees, under a policy that only lets deposit operations through.
  • TypeScript
  • Next.js 16
  • React 19
  • Hono
  • Zod
  • Stellar
  • Soroban
  • Passkeys
  • Vitest

Data Voyagers

Winning project · Alenia Consulting hackathon 2026

An agent that reads support tickets from OpenProject, works out which of three use cases each one belongs to, sets its priority and resolves it: it generates and runs the SQL, comments on the ticket, changes its status and records the resolution email. No action with side effects runs until a person approves it. Four of us built it with LangChain and LangGraph, deployed as a Databricks App; my part was the agent.

Learn more
  • Graph nodes never call the services directly: they go through a tool layer that rejects any action without a recorded approval.
  • To resolve a ticket it looks up similar closed tickets and uses their resolution as a precedent.
  • The queue runs in priority order and token usage is tracked per ticket, because cost counted in the judging.
  • Python
  • LangChain
  • LangGraph
  • Databricks
  • Streamlit
  • OpenProject API
  • SQL
  • pytest
  • mypy

Prisma

Hack4Edu 2026 · ProFuturo

A classroom tool that finds out, through short games, what each child is good at, and gives the teacher a report without labels: every figure says where it comes from and which games back it up. Built for schools with few resources: it works offline, with the pupils tablets or phones talking to the teacher laptop over the classroom network. A hackathon prototype built by a team of four.

Learn more
  • Six games with nothing to read, just icons and audio, covering logic, problem solving and creativity. Based on what it already knows about each child, the engine picks the games and difficulty for the next session.
  • Levels come from a public formula, with no language models and no training on data from other children, and each pupil is only compared with themselves. A test fails if any text contains a word that labels the child.
  • No runtime dependencies: Node 24 with its built-in SQLite, packaged as a single Windows executable.
  • JavaScript
  • Node 24
  • SQLite
  • Offline-first
  • node:test

AI Master's

Hate Speech Detector

Completed

Flags hate speech in YouTube comments. Instead of a yes or no, it returns a probability and a recommendation (allow, review or hide), and a person makes the final call. It can analyse a whole video from its URL, follow comments live, and has a browser extension. The model is an ensemble of two logistic regressions and a Naive Bayes, with a cross-validated macro F1 of 0.71 on a small dataset of about 1000 comments.

Learn more
  • The figures come from cross-validation because the test set has only 28 positives. Compared with the baseline, the ensemble improves ranking (PR-AUC +1.3 points) and leaves F1 unchanged: a real but modest gain.
  • Optuna tuned the hyperparameters with overfitting as a constraint; without it, it picked an ensemble with a 17-point gap between training and validation.
  • 91% of the hate in the dataset is racism about a single event, so it does not generalise to other kinds of hate. That is why review is the main recommendation, and hide is only suggested at high probability.
  • It never stores or returns author names, to comply with GDPR.
  • Python
  • scikit-learn
  • spaCy
  • NLTK
  • Optuna
  • FastAPI
  • Streamlit
  • Docker

Medallion Manager

Completed

Kafka sends each person in fragments (personal, banking, location and professional data) and none of them carries an ID. Medallion Manager rebuilds the whole record with a medallion architecture: MongoDB keeps the raw message (Bronze), Redis buffers fragments by match key with a TTL, and PostgreSQL holds the consolidated record (Gold). The reconciliation engine only merges what is unambiguous, such as the same passport with a similar name, or the identical name from sources that complement each other. Anything doubtful goes to human review, so two people with the same name do not end up merged into one.

  • Python
  • Kafka
  • MongoDB
  • Redis
  • PostgreSQL 16
  • SQLAlchemy 2
  • Pydantic 2
  • FastAPI
  • Streamlit
  • Prometheus
  • Docker

ClaimVox

Completed

Sorts written financial complaints (CFPB, 1.9M rows) into eleven product families. The model suggests and a person decides: human review is mandatory even when confidence is high. It was a team project and my part was the backend: the FastAPI foundation, JWT auth, Docker packaging, feedback stored in PostgreSQL under least privilege, and the AWS deploy.

  • Python
  • FastAPI
  • scikit-learn
  • PostgreSQL
  • Docker
  • AWS
  • JWT

Housing Price Predictor

Completed

Predicts the sale price of a home in Madrid with XGBoost, trained on about 94,000 Idealista listings (R² of 0.949, RMSE of €72,729). Around the model sits an MLOps pipeline: hyperparameters tuned with Optuna, experiments tracked in MLflow, weekly data drift reports with Evidently, and an endpoint to retrain on new data. A new model only replaces the current one if it improves R².

Learn more
  • Location goes in as the average price per m² of its area, on a 20 × 20 grid over the map.
  • Deployed on Render with Docker and CI/CD on GitHub Actions; drift alerts arrive on Slack. The demo may take a few seconds to wake up.
  • Python
  • XGBoost
  • Optuna
  • MLflow
  • Evidently AI
  • FastAPI
  • Docker
  • GitHub Actions
  • Render

LLM Content Generator

A proof of concept that writes ready-to-post copy for a blog, X, Instagram or LinkedIn, based on the topic, audience, tone and length you pick. It uses prompt engineering with LangChain and a Streamlit interface, and costs nothing to run: it works on the free tier of Groq or locally with Ollama.

  • Python
  • LangChain
  • Streamlit
  • Groq
  • Ollama

42 School

A-Maze-ing

Completed

A maze generator and solver in Python that you can also play. One pipeline produces two kinds of board: a perfect maze (a spanning tree, with exactly one route between any two cells) and a braided Pac-Man board with no dead ends. Algorithms are swapped through a Strategy registry and the generator installs as a wheel. The web layer is kept separate from the generator: FastAPI with SSE streaming, a canvas you move through with the arrow keys, and a chat with tool-calling that drives it in plain language. It has 300+ tests, run with a single make qa.

  • Python 3.12
  • FastAPI
  • SSE
  • Pydantic
  • pytest
  • mypy
  • LLM tool-calling
  • Docker
42 School projects

libft

Custom C standard library: string, memory, and linked list utilities from scratch.

  • C

ft_printf

Custom printf with variadic functions: handles %d, %s, %p, %x and more.

  • C

push_swap

Several sorting strategies (selection, chunk and radix), picking one based on how disordered the stack is.

  • C

so_long

2D top-down game using MLX42.

  • C
  • MLX42

minitalk

Inter-process communication using UNIX signals.

  • C
  • UNIX

pipex

Shell pipe simulation between processes.

  • C
  • UNIX

get_next_line

Reads lines from any file descriptor with a single static buffer.

  • C

Python Module 00

Python fundamentals: types, functions and control flow.

  • Python

Python Module 01

OOP: classes, inheritance, encapsulation, and polymorphism.

  • Python

Python Module 02

Exception handling: try/except, raise, custom errors, and finally.

  • Python

Python Module 03

Collections: lists, tuples, sets, dicts, generators, and comprehensions.

  • Python

Python Module 04

File I/O: read, write, streams, and context managers.

  • Python

Python Module 05

Abstract classes, polymorphic dispatch, and data pipelines.

  • Python

Python Module 06

Imports: packages, absolute/relative paths, and circular dependencies.

  • Python

Python Module 07

Design patterns: abstract factory, mixins, and strategy.

  • Python

Python Module 08

Virtual environments, pip/Poetry, and environment variables.

  • Python
  • NumPy
  • Pandas

Python Module 09

Pydantic v2: BaseModel, validators, and nested models.

  • Python
  • Pydantic

Python Module 10

Functional programming: lambdas, closures, decorators, and functools.

  • Python

Skills

Systems

  • C
  • Java
  • Python
  • Bash
  • Linux
  • Git

Backend & Data

  • Spring Boot
  • FastAPI
  • PostgreSQL
  • Redis
  • Docker
  • AWS
  • RabbitMQ
  • Flyway

ML / AI

  • XGBoost
  • MLflow
  • Optuna
  • Evidently AI
  • RAG
  • LLMs

AI Tooling & Workflow

  • Spec-Driven Development
  • Harness Engineering
  • MCP
  • Agentic Coding

Frontend

  • Astro
  • React
  • React Native
  • TypeScript
  • Tailwind
  • Next.js
  • Firebase

Contact

From idea to production.