Building useful products. Making complex systems easier to use. Contributing upstream.
Software Engineer at HP Barcelona
3DaaS — 3D Printers as a Service · Since June 2026
I build AI products, cloud platforms and open-source tools. My background spans cloud research at Cloudlab (URV), serverless computing on MareNostrum 5, and applied AI. Today, my main personal projects are Klipso and Invade.io, which I continue to develop alongside my professional work.
MSc in AI & Cybersecurity · Speaker at PyConES & EuroSciPy · IEEE-published research · Barcelona 🇪🇸
🎬 Klipso — AI-Powered Video Editing
AI-powered video editing for TikTok, Instagram Reels and YouTube Shorts. I built and continue to improve the full-stack product: highlight selection, speaker-aware reframing, captioned GPU rendering, a creator dashboard, credits and Stripe billing.
| Users | Clips generated | Video generated |
|---|---|---|
| 100+ | 1,000+ | 14+ hours |
Product milestones · October 2026.
The pipeline: Captions / Whisper → Gemini highlights → Speaker-aware framing → FFmpeg → Download
Recent engineering work:
- 4 → 1 encoder passes per clip, combining crop, upscale, subtitles and outro.
- 188.8 → 25.1 seconds for worker dependency installation in the documented T4 comparison.
- 548.4 → 276.6 seconds of worker elapsed time in a same-video, captions-enabled comparison; not a guaranteed total wait time.
- Recoverable job processing, idempotent status updates, automatic failure refunds and refresh-safe tracking.
React TypeScript Python Whisper Gemini MediaPipe FFmpeg Firebase Cloudflare R2 Stripe
🎮 Invade.io — Mobile Strategy & Conquest
A mobile-first territorial strategy game built with Godot 4. Capture bases, move troops and upgrade your army across a 30-level campaign, with continent-specific rules, bosses, daily challenges and free-conquest expeditions.
I develop the gameplay, progression systems and mobile experience, including local save recovery, accessibility settings and Android/iOS build workflows. Actively in development; cloud saves and online leaderboards are not part of the current public version.
Godot 4 GDScript Mobile Real-time strategy Game systems
📄 LiteRAG: Cost-Efficient Graph-Based Retrieval-Augmented Generation — arXiv
I designed and implemented a GraphRAG framework that replaces LLM-guided graph navigation with algorithms, exploring lower-cost, multi-hop retrieval.
Reported benchmark: 1.42s vs 142.04s on a 1,280-document corpus. Paper submitted to EMNLP 2026.
LiteRAG's navigation avoids LLM calls; answer generation and evaluation may still use models. Benchmark figures describe the evaluated setup, not a universal complexity or performance guarantee.
Python PyTorch Neo4j NetworkX Sentence-Transformers
🚀 PyRun Cloud — AI-Assisted Cloud Workspace
I led end-to-end development of a platform that helps researchers run distributed Python workloads on AWS and IBM Cloud: web workspace, automated runtime builds, cloud orchestration, execution dashboards and AI-assisted operations.
React TypeScript Python AWS IBM Code Engine Docker MCP
⚙️ Lithops — Multi-Cloud & HPC Computing
Open-source Python framework for parallel execution across cloud and HPC environments. My contributions include Kubernetes/RabbitMQ work queues, the Singularity HPC backend and execution reliability fixes, including work on MareNostrum 5.
📄 Enhancing HPC with Serverless Computing: Lithops on MareNostrum5 — IEEE ICNP 2024
Python Kubernetes Singularity RabbitMQ AWS Dask
🔊 Soundless — Citizen Science, Noise & Sleep
I led development across the Android app, Fitbit integration, cloud ingestion and analysis workflow, creating research infrastructure to study associations between environmental noise and sleep.
Kotlin Fitbit API Firebase GCP BigQuery Python
I built an LSTM-based race-ranking pipeline with lap-by-lap feature engineering and FastF1 integration, exploring how race sequences can inform predictions.
Python TensorFlow FastF1 Pandas
| Project | What it does |
|---|---|
| MedGround | Medical-document RAG prototype with cited answers, hybrid retrieval and LLM-as-a-judge evaluation. Built for document exploration, not clinical decision-making. |
| Research to AI Video | Prepares manually created NotebookLM videos for YouTube with Gemini-drafted metadata, human review and approved API uploads. |
| Spotify to Apple Music | Transfers playlists from Spotify to Apple Music with track matching, GitHub Actions and Telegram notifications. Builds on earlier open-source work credited in the repository. |
Specific changes, not just forked repositories:
| Project | Contribution | Status |
|---|---|---|
| Lithops | Kubernetes/RabbitMQ backend · Singularity backend | Merged |
| Dask CloudProvider | IBM Cloud support · Registry support and custom commands | Merged |
| Ollama | WebP image support in the desktop app | Merged |
| Gemini CLI | Fix URL.parse compatibility on Node.js <22 |
Merged |
| GitHub MCP Server | Strict validation of unknown toolsets | Open proposal |
| Rechunker | Lithops executor | Open proposal |
Contribution status checked in October 2026. Linked PRs remain the source of truth.
- PyConES 2025 · Sevilla — Multi-Cloud Serverless Parallel Programming in Python (Lithops).
- EuroSciPy 2025 · Kraków — Processing Cloud-Optimized Data in Python (Dataplug).
💻 Technologies & Engineering Toolkit
| Area | Technologies used across my projects |
|---|---|
| AI & retrieval | PyTorch · TensorFlow · Gemini · Whisper · MediaPipe · ChromaDB · Neo4j · NetworkX |
| Cloud & systems | AWS · GCP · IBM Cloud · Kubernetes · Docker · Singularity · RabbitMQ · Lithops · Dask |
| Product & interfaces | React · TypeScript · Python · Kotlin · Godot/GDScript · FFmpeg · Firebase · Stripe |
| Engineering practice | API integration · CI/CD · Observability · Performance profiling · Automated testing · Accessibility |
Explore the engineering behind these projects → macarronesc.github.io