Verdicts that changed a decision.
Thirteen client missions completed since 2025. The cases below are the ones I can describe in enough detail to be useful; the facts are from the engagements, and client names are withheld.
A trading strategy stopped before it risked real capital.
- Forex, triangular mispricing
- Historical backtest on broker data
- Live forward test on a demo account
Challenge
The client wanted an automated system that trades triangular mispricings in currency pairs, and needed to know whether the strategy had a real, deployable edge before funding a live account.
Approach
The system was built and tested in two stages: a historical backtest on the broker's own data, then a forward test running live on a demo account. When the first backtest looked promising, it was corrected for the costs and timing effects that flatter paper results.
Result
The corrected backtest showed an edge too small to justify deploying real money. The client stopped at that point, with the evidence and the code in hand, instead of funding a live account. The client is available as a reference.
A magnet design shown to miss its target before a prototype was built.
- Baseline 0.470 T computed vs 0.472 T analytical
- Best-case bound 1.149 T against a 1.2 T target
- Four hours, report plus Python script
Challenge
A bench rig for high-gradient magnetic separation needed 1.2 T in the working gap. The client wanted to know whether the proposed magnet and yoke could reach it before building anything.
Approach
First a baseline that could be trusted: the bare-magnet field was computed numerically (0.470 T) and analytically (0.472 T), and the client's own team confirmed the agreement. A finite-element approach was tried and abandoned because it smeared the discrete magnet array into something physically unrepresentative. A reluctance-network model then gave an upper bound for the field with zero leakage, the most optimistic case the design could achieve.
Result
The zero-leak ceiling came out at 1.149 T, below the 1.2 T target, so the design could not reach it even in the best case. The phase was delivered in four hours as a written report with the Python script that reproduces every number, and the directions with the most leverage for a redesign, a thinner gap wall and a closed back-iron return, were flagged for the next phase.
A medical question-answering system grounded in the client's own documents.
- Built on MedGemma
- Retrieval over the client's document set
- Delivered as a working pipeline with code
Challenge
The client needed a medical language model that answers from a specific body of documents rather than from whatever the base model remembers, with the answers traceable to the source passages.
Approach
A retrieval-augmented pipeline on MedGemma, Google's open medical model family: the client's documents are indexed, the relevant passages are retrieved for each question, and the model generates its answer from them.
Result
A working pipeline delivered with its code, so the client can run it on new documents and extend it. The client is available as a reference.
A measurement team trained on its own analyzer in five days.
- Keysight B1500A and EasyEXPERT
- Five days, four hours a day, on site
- Diodes, BJTs, MOSFETs
Challenge
A university team had a Keysight B1500A semiconductor parameter analyzer and needed to use it with confidence on their own devices.
Approach
A five-day course of four hours a day, delivered in person in Kuwait in autumn 2026, built around the team's own instrument and the measurements they needed: characterisation of diodes, bipolar transistors and MOSFETs with EasyEXPERT.
Result
A team that runs its own measurements, with course material it keeps. Tailored training of this kind is available as an add-on to any engagement.
The other missions, by field.
The thirteen missions since 2025 span quantum computing, materials science, semiconductor devices, biomedical and clinical study design, astrophysics, mathematical validation of models and code, trading systems, and LLM and retrieval pipelines. Details of any of them are shared on request, with the client's permission.
Start with one question.
Send a short description of the technology and the decision it informs. I reply personally, usually within one working day, and a first call is twenty minutes with no commitment on either side.