Tito Nicola
Drugman

A retrieval-augmented generation pipeline for code, built to run entirely locally. Hybrid retrieval — sparse BM25 fused with dense CodeBERT embeddings via reciprocal-rank fusion — draws from a ~7,600-snippet knowledge base and conditions a 4-bit-quantised CodeLlama-7B-Instruct. Measured with CodeBLEU: 0.186 → 0.257, a 38% relative gain over the no-retrieval baseline.

Deploying a transformer where there is barely room for one. An encoder-only ~4.56M-parameter model, quantised INT8 per-channel and compiled for the Neural-ART NPU on the STM32N6, running at 73 ms/token cold-start inside a 7.16 MB flash footprint. Built with Patrizio Acquadro.

A vision-language browsing agent benchmarked on WebVoyager. Ran Qwen3-VL (8B and 30B) locally through Ollama and DeepSeek-V3 via OpenRouter, against a GPT-4o baseline at 62.7% task success. Co-developed a grayscale Set-of-Mark prompting variant (with Montico) and added a ResNet-18 page-type classifier that cut agent latency by 17%.

Bachelor's thesis at STMicroelectronics: a hyperparameter-optimisation layer extending Keras Tuner to respect hard ROM, RAM and MACC budgets on STM32 targets — so the search only ever proposes models that will actually fit and run. Cost estimators validated on three MLPerf Tiny tasks; end-to-end HPO demonstrated on two. Supervised by Davide Denaro.

GenHack 2025 (École Polytechnique · BNP Paribas · Kayrros). A Random Forest correcting ERA5's urban-heat-island temperature bias — roughly 93% reduction in systematic bias and 35% in RMSE. The dominant predictors turned out to be elevation, latitude and longitude rather than urban fraction, which reframed how the team read the error.