Email triage and draft-reply agent
It sorts every new message into reply or no reply, and writes a draft for the ones that need one. The model runs on your own computer, so no email leaves the building.
// selected work
Select any project to read the problem, the solution and the hours saved.
It sorts every new message into reply or no reply, and writes a draft for the ones that need one. The model runs on your own computer, so no email leaves the building.
A PyTorch network that predicts the best aerofoil shape for a set of flight conditions. Move the sliders to see the shape change.
Built for a leading Canadian university. It replaced hand scraping: the program reads the records through the API and writes correct CSV files.
An API layer that connects Raspberry Pi hardware to modern model and cloud services, so a small board can do new work.
// let’s build
Tell me the slow job. I tell you how I would automate it. I reply in one working day.
Featured project
The agent sorts every new message into reply or no reply and drafts an answer for the ones that need one. The model runs on your own computer, so no email content leaves the business.
A busy inbox takes an hour or more every morning, and the owner of the business is usually the one who reads it.
The agent runs before the working day starts and does three steps for each new message.
Nothing is sent automatically. You read the draft, change what you want, and press send. If the agent is not sure, it flags the message and writes no draft.
Private by design. The language model runs locally on the machine. No email content goes to an outside service, which keeps client correspondence inside the building.
Five hours a week for one person, and a shorter reply time, because the answer is ready before the day starts.
Model development
A trained PyTorch network that predicts the best aerofoil shape for a given set of flight conditions, and returns the top profiles ready to plot.
An engineer must choose an aerofoil for each new set of flight conditions.
I built the dataset from the start, then designed and trained a network on it with a CUDA GPU.
The sliders on the card show the same idea: the shape changes as the conditions change.
The result is also repeatable: the same conditions always give the same answer.
Tool and API integration
Built for a leading Canadian university. The pipeline replaced hand collection: it reads the records through the API and writes correct CSV files, with a check on every field.
The team collected the records by hand from the web interface, page by page.
A Python pipeline that does the full job in one command.
Two weeks of collection time a year goes back to the team. Accuracy is the second gain: the program reports a bad field, so nobody checks the file by eye.
Hardware integration
An API layer that connects Raspberry Pi hardware to modern model and cloud services, so a small low-power board can do work that the platform could not do before.
A Raspberry Pi is small, cheap and low-power, and it reads sensors and controls devices well.
A bridge between the hardware and the API. The board collects the data and acts on the result.
This project does not save hours. It adds a new ability: a £40 board can use the same models as a large computer, and it runs where a large computer cannot go — a workshop, a vehicle, a field cabinet or a shop counter.