Engineering lead.
I build machine learning systems that run in production.
Now
Engineering lead
Stealth startup in aerospace, defence and robotics
Founder
Previously co-founded Katalo
Before
Machine learning
Renewcast, tem. and AmpX
Based in
Prague
Remote with UK and EU teams
7,714 contributions in the last year
github.com/dean-sh7,714 contributions in the last year across 368 days. Busiest day 2026-02-19 with 123.
Selected work
Built to workoutside the notebook.
Weather forecasts can't see what a plant did an hour ago. This model can.
A small network reads each plant's latest readings and corrects the next four hours. One model covers the whole solar fleet.
21.9%
Median cut in solar error over the first two hours
Read the case study
An AI that applies to jobs for you, only where you said yes
Otty finds roles and applies for you over WhatsApp. The model decides what fits. Code checks your rules before anything goes out.
5×
Lower cost per agent turn
~0.9 s
Time to first response, down from 6.3 s
Read the case study
Cutting solar forecast error from 15% to 6% in five months
Most of a solar plant's output comes down to sun angle, panel layout and temperature. Physics handles that part. I trained the model only on what physics gets wrong.
15.3% → 6.2%
Portfolio forecast error, May to October 2025
Read the case study
One forecasting system for portfolios that all look different
One client had 1,189 plants behind 1,104 meters. Another needed forecasts for individual zones. A third had meter data for only some sites. Each one used to mean a new pipeline.
4
Client portfolios with different layouts
Forecast vs metered energy, by month Read the case study
AI can stage a living room. It shouldn't move the walls.
Image models are good at furniture and bad at architecture. I built the pipeline that checked every edit and repaired the ones that changed the room.
95%
Precision against human reviewers
How the LLM judge decided Read the case study
Earlier work
- Pricing energy contracts against the risk of the whole portfolio95%Less time to price a tender
- Matching small businesses with local renewables to skip £50/MWh in levies£50/MWhLevies avoided per matched MWh
- Forecasting load, solar and prices for hundreds of sites at once>30%Lower MAPE than the benchmark
- One model contract, so every forecast ships the same way<1 dayDeploy prep, down from 4-5 days
- Catching robot failures on the line before they happen>35%Less unplanned downtime
“Dean has a real talent for clarity. The way he presents information makes even complex, data-heavy content easy to follow. His communication style is clear, inclusive, and well considered.”