Tommi Bimbato

MSc Data Science, Machine Learning and Statistics — Uppsala University, Sweden

GitHub: github.com/tbimbato
ORCID: 0009-0003-0667-3230

About Me

MSc in Data Science at Uppsala. Architecture degree at IUAV Venice, five years of practice before switching fields.

Interested in how data science can model systems that have a physical reality, medical and health data, acoustic spaces, complex environments. Currently working on acoustic scene classification and machine learning and signal processing. Self-taught in DSP: Max/MSP, gen~, audio modelling.

Still figuring out where these things connect.

Outside: trail running, hiking, gardening.

Education

| Uppsala University, SwedenMSc Data Science, Machine Learning and Statistics (2025 - 2027)
Machine learning · statistics · time series analysis · deep learning · data engineering (Hadoop, Spark) · combinatorial optimization (MiniZinc) · software engineering · scientific computing

| University of Verona, ItalyCoursework in Computer Science and Data Science (2023 - 2025)
Statistical learning · probability · databases · mathematical logic · computer architecture & assembly · C · SystemVerilog · FPGA prototyping · algorithms and data structures

| IUAV University of Venice, ItalyBA Architecture (2020)

Experience

| STEM teacher, FabSchool (2024 - 2025)
Teaching in STEM subjects, within outreach initiatives dedicated to STEM education and public science communication.

| Architect (2018 - 2023)
Architectural practice, across more than twenty projects.

Projects

| 2026  Acoustic Space Classification: Classical ML vs Neural Baselines — Controlled sim-to-real study on room-type recognition from impulse responses (IRs). Six interpretable acoustic parameters (RT60, EDT, C80, D50, Ts, DRR) with classical classifiers, benchmarked against CNNs and a pretrained ResNet18 on the identical protocol: trained on simulated rooms (pyroomacoustics), tested on real held-out rooms from BUT ReverbDB, AIR and ACE. The physical features transfer more robustly than the learned representation (accuracy drop from simulation to real rooms: 0.37 vs 0.52), and the gap widens as the simulator is made more physically realistic. Includes a published 5,000-RIR synthetic dataset with a datasheet.
| [Repository] | [Dataset DOI]

| 2026  MCP RIR: Room Impulse Responses from Natural Language — Model Context Protocol server that turns a spoken description of a space into a simulated room impulse response. The calling model maps free text in any language onto room geometry, absorption coefficients and mic/source placement; the server runs the image-source simulation (pyroomacoustics) and writes a stereo 48 kHz WAV ready for any convolution reverb. Unspecified parameters are randomised per call, so repeated requests give different takes of the same room. Intended for matching a location no longer accessible when no IR was captured, and for building reverbs of rooms that could not be built.
| [Repository]

| 2026  Helsinki Urban Noise Analysis: Hourly LAeq Time-Series Modelling — Comprehensive SARIMA analysis of hourly equivalent sound pressure level (LAeq) measurements from a Helsinki urban sensor. Stationarity testing (ADF/KPSS), seasonal differencing, ACF/PACF-guided model identification. Performed model diagnostics (residual analysis, Ljung-Box test) and forecasting evaluation.
| [Repository]

| 2025  Diabetes Classifier: EDA, Benchmarking & Interactive Prediction — End-to-end ML pipeline on a clinical diabetes dataset (blood markers: HbA1c, BMI, lipid panel, renal indicators). Physiologically-motivated outlier detection, unit inconsistency discovery (VLDL mg/dL vs mmol/L), benchmark of four classifiers on balanced and imbalanced splits. Streamlit dashboard with interactive patient input, confidence scoring, and cross-model comparison.
| [Repository]

Awards

| Nov 2025  — 1st place, AWS Immersion Day Hackathon — Stockholm
| Nov 2025  — 3rd place, Lovable x UUAIS Hackathon Night — Uppsala

Research Output

| 2026  Bimbato, T. ASC-26: A Synthetic Room Impulse Response Dataset for Room-Type Classification [Dataset]. Zenodo.
5,000 labelled room impulse responses across ten room types, simulated with pyroomacoustics and documented with a datasheet (taxonomy, generation procedure, per-class acoustic statistics, stated limitations). CC BY 4.0.
| [10.5281/zenodo.21771185]

Writing

| 2026  Deep Learning — Informal Notes (PDF) — Informal notes on deep learning.
| [PDF Link]

| 2024  Probability Theory Handbook & Formula Sheet — Reference handbook and formula sheets.
| [Repository]
| [Handbook PDF]

Earlier Work

| 2018  — Mentioned in IQD n. 53 (Singolarità) and cited in Lampedusa: La Cattedrale di Solomon — Vol. II as student-contributor to a project with Arch. Renato Rizzi.
| [ResearchGate]

| 2017  — Credited as student contributor in F. Cacciatore, Rubble or ruins? (Università Iuav di Venezia, series Syria. The making of the future), ISBN 8899243255.

| 2015  — Co-designed Nothing Personal — exhibition at IUAV's Ex Cotonificio combining spatial layout and immersive audiovisual experience. Extended beyond its initial run due to public and academic reception. Team of six.