Turning supplier emails into structured quotes
An aviation spare-parts company received supplier quotes as free-form emails. I built a system that reads their mailboxes and turns each real quote into a structured record. I built and integrated it alone.
Problem
Prices and part details were buried in PDF, Word, Excel or ZIP attachments, so buyers had to open, read and re-type every quote.
Approach
The system reads the mailboxes, recognises real quotes and extracts part numbers, condition codes, prices and certificates into the procurement platform already in use. Follow-up replies update the same record, failures are retried, and buyers can review and correct every result.
Result
- In daily use
- Much improved how the team processes and searches supplier quotes
Relevant if your team re-types quotes, orders, invoices or shipping documents.
Deciding on board which satellite images are worth sending
For a company building embedded space systems, I designed a filtering system that decides which images to keep and send down.
Problem
A satellite has limited compute and bandwidth, so it cannot send every image. The choice has to be made on board, and accurately enough to be worth the energy.
Approach
The system balances accuracy against the compute and bandwidth available, and can learn new targets from a few examples. The same models run on servers or on embedded hardware.
Result
- A working version is complete and is being improved
- A much better balance between accuracy and the data and energy budget
Relevant if AI has to make decisions on limited hardware, such as cameras on a production line.
RetroBench: a test AI vendors can't game
RetroBench tests whether AI coding agents can rebuild source code that compiles byte for byte to original 1990s game code.
retrobench.ai
Problem
Public benchmark tasks end up in training data, so a high score can mean the model remembers the answer and not that it can do the work.
Approach
The tasks ask for source code that compiles to exactly the original bytes. No model can have memorised them, and the result is checked by the compiler, not by opinion.
Result
- 15 models from the major labs tested
- Scores range from about 27% to 98%
Relevant if you are choosing between AI models or vendors: I build the same kind of test from your own cases.
Demand and supply modelling across sports for a national programme
For a national programme, delivered with a UK partner, I built models that answer a planning question from the data that could actually be found.
Problem
The programme needed estimates of supply for different sports, both overall and by region. The inputs were demographic and statistical data, survey data and research findings, and none of them covered the question completely.
Approach
I built the models to fit the data that was available, with assumptions that can be explained in plain words and changed easily.
Result
- Working models delivered
- Assumptions are explicit and easy to change
Relevant if you have a planning question and partial data, but no one to model it.
On-premises document search for data-sensitive environments
- Technology
- Documents
- On-premises
A document search and AI access system for organisations whose data cannot leave the building.
Problem
AI systems that use a standard API process the data outside the organisation. That is not acceptable when documents are confidential or contain personal data.
Approach
I built and delivered a system that runs the AI models and the indexing models locally, and provides semantic search and AI access over the documents.
Result
- People find the documents they are looking for
- Documents feed the local AI models, so the information in them is used more effectively
Relevant if your documents cannot be sent to an outside AI service.
Predictive maintenance in manufacturing
- Manufacturing
- Predictive maintenance
- Sensor data
An example of a first project for a manufacturer that records machine data and still learns about failures when the line stops.
Problem
Unplanned stops are expensive, but so is replacing parts too early. Whether the sensor data contains a usable early signal is not known until it is tested on past failures.
Approach
I would take the failure and maintenance records you already have, build a test set from past stops, and measure how early and how reliably a model would have warned.
What a first test would measure
- How many past failures would have been flagged in advance
- How much warning time there would have been
- How many warnings would have been false