WooAh Choi

MSc AI at Queen Mary University of London, with three years of industry experience as a Software Engineer at Konan Technology building enterprise search and NLP systems. My coursework spans Machine Learning, Neural Networks and NLP, Information Retrieval, Reinforcement Learning, and Conversational Agents — building depth in both theory and implementation.

Most recently I built ResearchAgentLab, an autonomous research agent, and then a control architecture for it after watching it silently substitute a different experiment for the one I had approved. Protocol Before Code freezes approved intent in a machine-checkable protocol, constrains how much the model is allowed to generate, and checks scientific conformance independently of software correctness. In holdout and fault-injection evaluations, 0 of 18 drifted protocols passed and 18 of 18 behaviour-changing trusted-component mutations were blocked before execution; with the content-integrity check neutralised, 15 of those same 18 escaped the remaining stack. A frozen 192-pair oversight experiment found that adversarial curation moved an overseer's target selection from 13.0% to 27.1% while explicit approval stayed at 100% in both conditions. Written up as a sole-authored technical report.

My dissertation is a mechanistic analysis of a PPO-trained causal Transformer policy in a partially observable MiniGrid T-maze. Its central finding was methodological: the aggregate success metric hid a failure mode, because an agent that always turns the same way scores around 50% while having learned nothing about the instruction cue. I designed a worst-case metric, Smin = min(S_key, S_ball), that exposes those one-sided policies, and used causal interventions to test control rather than infer it from correlation. Difference-vector attribution assigned 33–58% of the retrieval difference to padding, yet exchanging the padding source reversed no decision while exchanging the real-transition source reversed both — correlation and causation pointing at different things on the same policy. That signature held in only three of six checkpoints, so equal behaviour did not imply equal mechanism. A parameter-matched LSTM also solved the task at Smin = 0.898, which moved the contribution from architecture comparison to the analysis and evaluation framework itself.

Third author on a short paper accepted to FinNLP 2026 at EMNLP, on whether relevance-ranked web evidence is actually sufficient to verify financial claims. I was one of the two first-pass annotators, labelling 500 of the 750 claim–passage pairs under a blinded four-class protocol, including the 250-pair double-annotated block behind the paper's Cohen's kappa. I also review for the same workshop, the annual workshop of the ACL Special Interest Group on Financial and Economic NLP.

As co-founder of GridFlow Trade (selected for QMUL QIncubator), I built the UK day-ahead price forecasting product on 65,000+ real half-hourly settlement periods. The work I would point at is not the model. Early accuracy looked strong and was wrong: training showed MAE £7.4 while the same model served in production sat at £21.9, because training and inference were not seeing the same features. I rebuilt it as a single point-in-time feature pipeline shared by both paths and verified they produce identical features. The honest figure afterwards is MAE £24.2/MWh against a £34.2 naive baseline, a 29% error reduction on a 90-day held-out window averaged over five seeds. Benchmarking Naive, Ridge, LightGBM, XGBoost and a bidirectional LSTM on one leak-free split showed the tree models and Ridge were indistinguishable within seed noise, so the gains came from feature and evaluation design rather than model choice. I also audited 3.7 years of stored market data against the Elexon API and found a collector date bug that had corrupted 1,634 price and 1,634 system-margin records on every British Summer Time day.

At Konan Technology, I delivered production search and NLP systems for the Supreme Court of Korea, reducing litigation search latency from 5+ minutes to 8 seconds, and shipped query intelligence features across four enterprise clients.

Kaggle: Top ~17% in Google AI4Code · Top ~28% in RSNA 2022 Cervical Spine Fracture Detection.