About

I’m Tino Trangia — a software engineer and data scientist working at the intersection of ML research and production systems.

I’m finishing my M.S. in Data Science at UC San Diego (2026), and previously earned a B.S. in Electrical Engineering and Computer Science from UC Berkeley. My research interests are interpretable ML, LLM reasoning and evaluation, AI alignment and safety, and applied NLP.

What I’ve worked on

Software engineering at FabricFeed (2023–2025). I designed and built a full-stack ERP application (Django, PostgreSQL, Next.js) end-to-end — from database schema and API design through CI/CD deployment — along with an inventory and analytics dashboard that surfaced recycling trends and cost inefficiencies across dozens of textile waste streams. I also prototyped an LLM-based report generation agent that turns natural language requests into structured plans with SQL query composition, cutting report generation from about an hour of manual work to under ten minutes.

Graduate research at UCSD. I’ve built evaluation frameworks comparing 64 LLM model pairs across hundreds of scenarios for scalable oversight research, investigated how steering chain-of-thought reasoning length affects accuracy and faithfulness, and replicated a NeurIPS 2025 data extraction attack that exploits pre- and post-unlearning checkpoints to boost extraction of memorized data by up to 60%.

Earlier work. As a research assistant at Berkeley’s Space Sciences Laboratory, I built a Python application for isotope mapping with neutron imaging data on 10+ GB datasets (reducing load times by 70%), and wrote automated testing and visualization scripts for space-borne sensors. I also spent a year and a half as an academic intern helping teach introductory data science to Berkeley undergrads.

Selected projects

• Database Reporting Agent — multi-agent text-to-SQL pipeline (LangGraph, Chainlit, Postgres) with schema resolution, query generation, guardrails, validation, caching, summarization, and evaluation over a 15+ table enterprise database. Uses RAG over an embedded vector store for few-shot retrieval and semantic layer mappings.
• Data Extraction From Fine-tuned LLMs — replication of a NeurIPS 2025 attack on exact unlearning; LoRA finetuning on Llama models to characterize memorization at inference.
• CoT Reasoning Length Steering — selective weight editing of short-reasoning attention heads, with follow-up analysis of how reasoning length affects chain-of-thought faithfulness.

Stack

Python, SQL, TypeScript, Go, Java, C++. PyTorch, HuggingFace, LangChain / LangGraph, scikit-learn. Django, FastAPI, Next.js, React, Postgres, Neo4j. AWS, Docker, Git, Claude Code.

Get in touch

Email ttrangia@ucsd.edu · GitHub @ttrangia · LinkedIn ttrangia.