About
Research Software, Energy Data and AI Evaluation
I am Cristian Dinu, a software engineer working across energy, data and research. I test and evaluate AI systems and build agent-based simulations alongside that work.
Work
My research software work at UCL supports the Smart Energy Research Lab (SERL) and Energy Demand Observatory and Laboratory (EDOL). I design and build data infrastructure that gives UK researchers access to detailed energy datasets while maintaining strong privacy protections.
SERL and EDOL securely collect and enrich household energy data from smart meters, environmental sensors, surveys and IoT devices. The engineering challenge is to make these data useful to researchers while meeting the requirements of data governance and smart meter regulations.
Previously, I co-founded Hypersay, a platform for online events and real-time presentation engagement. Earlier in my career, I worked as a Risk Manager at IBM. Those experiences continue to inform how I build software: understand the problem, consider what could go wrong, and make the result useful to the people working with it.
Current interests
I test and evaluate software built on non-deterministic components, particularly LLMs and agents. The questions I keep coming back to: what counts as good enough, how do you investigate a failure, and how do you tell whether a change actually improved the system?
I use Inspect AI to test tool use and to see how language models respond to unreliable information. I build repeatable tests and write up the findings so they are useful: what happened, under which conditions, what needs changing, and what the tests have yet to establish — the same discipline I learned doing risk management.
Model selection is part of the same question. A smaller, cheaper model may be adequate for a particular task, but that needs evidence, not intuition. I compare models on reliability, cost, and the consequences of getting something wrong.
I also build agent-based simulations, including with Google DeepMind’s Concordia, to see how individual decisions and interactions produce wider patterns of behaviour, and how far a simulated environment can test an idea before it meets the real world.
This builds on earlier work with natural-language interfaces and the Model Context Protocol (MCP): connecting models to tools makes them more useful, and gives me more behaviour to test.
Evaluation itself is the new part. The underlying skills are not: I have tested software for years, worked as a risk manager, run privacy-governed research data systems, and co-founded a company. Engineer, risk manager, research-data lead, founder, and someone who reads the humanities — that combination is uncommon, and every piece of it is checkable against the work above.
Connect
You can connect with me on LinkedIn. Views on this site are my own.