I work on trustworthy AI: evaluating machine-learning systems, finding out where and why they fail, and explaining those failures with interpretable models. I build this as tested, reproducible research software, and I report what the evidence does not show as clearly as what it does.
MSc in Computer Engineering, University of Genoa (2026). I was a Research Assistant (MSc thesis research) at CNR-IEIIT in the EU Horizon Europe project REXASI-PRO, working on the reliability of neural trajectory predictors for smart-wheelchair navigation.
Research interests: failure analysis of learned components in safety-critical systems · ML evaluation (black-box evaluation, risk maps, held-out checks, confidence intervals, honest baselines) · interpretable models as failure rules · safety filters for learned navigation · reproducible research software · evidence-level evaluation of retrieval-augmented generation.
Each card links to its repository; the images are real outputs of those projects.
|
SafeTraj-Experiments · MSc thesis Black-box evaluation of five pretrained neural trajectory predictors for a smart wheelchair (REXASI-PRO): 200 sampled commands × 3 goals per model. Strict success ranged from 25.3 % to 99.3 % under identical conditions; failures concentrate at initial orientations near ±π, and depth-4 decision trees turn them into explicit rules. 42 checks recompute the thesis numbers from the published files. |
trajectory-failure-analysis · preprint The same approach on real pedestrian trajectories (ETH dataset, 719 trajectories, two scenes) for a damped constant-velocity predictor. All 16 preprint values are reproduced from pinned raw data. Additional analyses show the limits: held-out tree accuracy does not exceed a constant baseline, and part of the positional signal in one scene is consistent with a horizon effect. Preprint, not peer-reviewed. |
|
SafeTraj-Prototype · toolkit The failure-analysis workflow as a reusable tool for any trajectory predictor: trajectory metrics and rule checks, input-novelty percentiles (reported separately from outcomes, because a typical input can still fail), risk maps and decision-tree failure rules. One implementation behind a Python library, CLI, FastAPI service and Streamlit dashboard, with 46 tests. Hosted API (version 2.0.0, Render free tier). |
SafeNav-RL · simulation PPO navigation agents with a discrete-time control-barrier-function speed filter, its proof and stated assumptions, and a seeded multi-seed benchmark. With the filter on there were no collisions in 3,600 test episodes; the cost is deadlock (expert success 94.6 % → 65.0 % when the filter is added). Training with the filter did not clearly help: 76.4 % vs 70.0 % with sampled actions, with overlapping intervals. |
TrustRAG: evidence localisation and failure attribution for RAG evaluation
A measurement study of what happens inside the right document: retrieving the correct document is not the same as ranking the passage that supports the answer first.
- Seven localisers (BM25, dense bi-encoders and cross-encoder rerankers; 22M–278M parameters) on QASPER and Natural Questions, with a retrieval decomposition on four corpora.
- An author-conducted annotation of 200 QASPER units: gating failure attribution on span coverage agrees with the labels more often than gating on document coverage (accuracy 0.700 vs 0.600). One annotator: agreement with a guided expert reading, not independent validation.
- A defect found in my own BM25 baseline led to withdrawing an earlier result; the correction is kept on record in the repository.
Offset-carrying pipeline, failure taxonomy, Wilson intervals, exact McNemar with Holm correction, cluster bootstrap; 504 tests. Software and annotation data on Zenodo: 10.5281/zenodo.22879358, 10.5281/zenodo.22879569.
GroundedAI: evidence-grounded document question answering
▶ Try the live demo (runs entirely in your browser; no install) · Repository · Release v0.1.0
A retrieval-augmented question-answering application that checks its own answers: every statement is compared with the passages the model received and comes back with a status, the exact quoted sentence and its page. Conflicting sources are shown side by side, and when the documents do not say, the answer is "not enough evidence".
- The live demo runs the unchanged FastAPI backend in the browser (Python compiled to WebAssembly): PDF extraction, hybrid retrieval, the prompt and the claim check, with an optional Qwen2.5-1.5B model on the visitor's GPU (WebGPU).
- With qwen2.5:1.5b through Ollama on a 56-question synthetic set, 29 of 41 answerable questions were answered correctly with every statement verified, and 8 of 10 questions that should be refused were refused. The checks are lexical and rule-based, and the labels were written by the author or by models.
FastAPI · LangChain · SQLite FTS5 or PostgreSQL + pgvector · React · Pyodide · WebLLM; a browser check of the live demo runs in CI.
- ML and evaluation: Python · PyTorch · scikit-learn · NumPy · pandas · SciPy · Gymnasium · Matplotlib
- LLM and retrieval: sentence-transformers · BM25 · cross-encoder rerankers · ChromaDB · LangChain · Ollama · WebLLM · OpenAI and Anthropic APIs
- Software: FastAPI · Streamlit · pytest · GitHub Actions · Docker · Prometheus · Git
- superstore-analysis: end-to-end BI workflow with Python, SQL and Power BI.
The banner's project map is a diagram of the repositories above; its motion is CSS-only, and a still version is shown when your system asks for reduced motion. The project images are figures and a screenshot from those repositories, the SafeNav-RL animation replays its stored benchmark trajectories, and the GroundedAI animation is a recording of its live demo. All are generated by tools/make_profile_v2_assets.py.