Anonymised to sector. Each one started with an operational problem, not a technology.
Healthcare R&D · Portugal
Clinical notes captured during the shift, and accurate enough to trust
Getting nurses' notes written down as the shift unfolds was only half the problem. The R&D team could not guarantee that what was written was clinically accurate.
I designed a language model working in the vocabulary of the CIPE nursing manual, paired with a clinical reviewer agent that checks each note and flags anything inconsistent for a person to review. The model drafts. The reviewer decides what needs a human.
ResultIn laboratory testing, more than 98% of ambiguities were correctly identified and sent for human review. Note-taking time fell by 45 minutes per user per shift.
Document-heavy operations
Document extraction that cost too many tokens to scale
Sending every document through a language model worked, and was far too expensive to run at volume.
I redesigned the pipeline so the model does as little as possible. A Python extractor reads each PDF, a lightweight classifier identifies it, and a script generated for each document provider pulls out the fields. The cost falls because of how the work is divided, not because of a cheaper model.
Result86% fewer tokens used, with similar extraction performance.
Non-profit sector
Funding calls found too late to answer, or not found at all
A group of NGOs needed to follow grant opportunities from international programmes down to the hyperlocal, and the calls surfaced too late to answer or never surfaced.
I built a crawling and classification system that gathers very different kinds of opportunity into one canonical database, then scores each against the organisation: how well it fits, how large the potential is and what pursuing it would cost.
ResultMore than 800 sources monitored, with a median time between a call being published and captured of under 24 hours.
Professional services
Month-end reporting that took 19 man-days
Every month, a team spent 19 man-days producing activity and status reports, then distilling them into a short executive overview and a decision matrix.
I designed a system that integrates the organisation's separate data sources into one canonical state-transition model of the business and derives the reports and the matrix from it. The reporting becomes a view of that model rather than something assembled by hand.
ResultFrom 19 man-days to 0.5, spent mostly reviewing the mismatches the system flags itself. It learns from that input, so the figure should keep falling.