Work
Case studies from live engagements, not illustrative examples.
Each of the engagements described below is live, named with the client's consent, and operating in production. A new case study is published only once there is something substantive to describe, and once the client has agreed to the particular form of disclosure. The aim is depth over breadth.
Internal research
Tools built in-studio, with the implementation published in full.
Research engagements undertaken inside the studio rather than for a client. The interactive interface for each is embedded on the case study page below, so the work can be evaluated directly rather than only described.
-
Internal research
London Property Finder
An interactive single-page tool that scores every current Rightmove London listing against a stacked machine-learning ensemble trained on 2.18 million Land Registry transactions. Stacked XGBoost with quantile-loss meta-learner; gap MAE £38,832 (-32% vs baseline) and 68.3% within ±10% of asking. Sold-nearby comparables, 80% prediction interval, filterable list and map.
Applied machine learningGeospatialQuantile regressionOpen the case study -
Internal research
London Auction Property Finder
A planned auction-specialist machine-learning model abandoned mid-build when the obvious training corpus (HM Land Registry's price-paid category B) was found to be a heterogeneous bucket of non-standard transactions whose median is +47% above standard. The case study documents the data audit, the resulting pivot to a flag-and-discount tool surfacing auction-eligible listings with a literature-derived 80-95% retail-discount band, and what auction-house data a v2 would require.
Data auditDisciplined model abandonmentApplied MLOpen the case study -
Internal research
London Property Valuation
A self-service valuation form that scores any user-described English property through the same V1.1 per-city ensemble that ranks live listings. Postcode-aware dispatch across twenty-one cities, k-nearest-neighbour imputation for missing input features, and an adaptive six-tier comparable-sales finder with HPI-adjusted prices. The case study documents the architecture trade-off between managed cloud hosting and Cloudflare-Tunnelled self-hosting, and the limits within which the tool can defensibly be used.
Self-service toolingPer-city dispatchSelf-hosted backendOpen the case study
In preparation
Further engagements awaiting the client's approval to disclose.
The entries below describe work already in production. Each case study will be published in full once the client has reviewed the proposed disclosure and agreed to its specific form.
-
Rud Pedersen Group
Sentiment intelligence engine
Sentiment extraction and signal classification at scale across a public-affairs consultancy's media and briefing corpus. A case study is in preparation and will be published once the client has approved disclosure.
Data engineeringNLPPublic affairs -
Rud Pedersen Group
Retrieval-augmented knowledge system
A retrieval-augmented generation pipeline over the firm's internal research and briefings library, with hybrid retrieval and reranking tuned to the document mix. A case study is in preparation.
Retrieval-augmented generationAzurePublic affairs -
UCL · Royal Free Hospital
Cell-spheroid segmentation
A U-Net segmentation model for high-throughput cell-spheroid microscopy, delivered as research engineering for a surgical biotechnology group. Validation intersection-over-union of 0.968 and Dice coefficient of 0.983. A case study is in preparation.
Computer visionPyTorchClinical research