Research Engineer · AI MSc @ University of Amsterdam · RL, world models, LLM quantization & efficient inference
Website · Email · LinkedIn · Google Scholar · CV
I'm Francesco, a research engineer and AI MSc student at the University of Amsterdam with a background in software engineering and machine learning research.
I like building things that sit between research and engineering: reproducible ML pipelines, model evaluation setups, world-model experiments, and tools that make results easier to inspect, explain, or deploy. Lately I'm most interested in efficient inference.
Before focusing more deeply on AI research, I worked as a full-stack developer, building production web platforms with React, Node.js, PostgreSQL, Docker, and Linux deployments.
GPTQ quantization of ALMA-13B-R to 8/4/3/2 bits, with memory and latency profiling on H100 across 10 WMT translation directions.
- 4-bit: 2.3× less peak VRAM (28.8 → 12.3 GiB) for −0.8 XCOMET-XXL.
- Profiling: traced slow quantized decoding to non-fused kernels that dequantize every weight matrix on each forward pass. Wrote a lossless repacker that moves 3-bit weights into the fused ExLlamaV2 4-bit layout, making generation 2.3–2.6× faster.
- Recovery: a 0.47 GiB LoRA adapter distilled from the fp16 model brings the collapsed 2-bit model from 22.7 to 88.1 XCOMET-XXL (94% of the gap closed) in 5.5 GiB of GPU memory.
A research project on hierarchical planning for goal-conditioned control using latent macro-actions, CEM/MPC planning, and a frozen low-level world model. NeurIPS 2026 PTA Workshop and WM@Booth 2026.
Layerwise probing of V-JEPA, VideoMAE, and LTX-Video representations to study whether pretrained video models encode intuitive-physics structure. NeurIPS 2026 World Models in Physical AI Workshop.
Reproduction and extension of harmful-content filtering pipelines for web-scale datasets using open-weight LLMs and moderation benchmarks. Under review at TMLR.
Machine learning pipeline for transcriptomics analysis, combining XGBoost, SHAP, differential expression analysis, and biological enrichment. First author, arXiv:2603.05572.
AI / ML
Software engineering
You can reach me at massafra32@gmail.com or connect with me on LinkedIn. More at firewtap.github.io.