The Atelier is where Ce’s project students are listed, together with what they built. It is called an atelier because the arrangement is an old one: students learn by making something, under a practitioner with firm opinions about how it should be made. It is virtual because there is no room; the work happens in repositories, shared drafts and video calls. It is of the alchemical kind because the proprietor has yet to produce any gold himself. The students have the better record, as the entries below show.

The list is far from complete. Any past student who finds this page and would like to appear on it need only email Ce at his usual address.

At the bench

  • Tong Zhao (PhD) works on learning causal structure from time series: which variables drive which others, with what delay, and in what order, recovered from the data alone by methods fast enough for high-dimensional series. She also co-wrote ResBench, a benchmark that judges FPGA designs written by language models by the hardware resources they consume as well as by whether they work. [pdf]

  • Duru Yağmur Akyol LinkedIn (MSc) built LLM4HLS-Agent, a controller that lets a language model repair and improve hardware designs written for high-level synthesis, within a fixed budget of model calls and tool runs, and that keeps only the designs it has verified. The work won 🥇 First Prize in Design Competition Track A at the International Conference on Field-Programmable Technology (FPT), 2026. [pdf] [code]

  • Jiajun Li (MSc) built a controller that decides, before each trade, how far a learned option-hedging policy should be trusted, using only the outcomes observed so far, and that falls back to the classical Black–Scholes hedge when the evidence turns against the policy. The paper, “Conformal Trust Control for Learned Sequential Policies under Regime Shift”, was accepted at the ACM International Conference on AI in Finance (ICAIF), 2026. [pdf]

  • Hannah O’Keeffe (MSc) built a learning pipeline that decides how much capital a bank should hold against credit losses, and that is trained on the quality of that decision instead of the accuracy of the forecast behind it. The pipeline passes through an estimate of default probability that a regulator can inspect, and comes with an audit framework that checks the estimate against supervisory validation criteria.

Out in the world

  • Max Stoddard LinkedIn (BEng, 2026) improved the agent-based model of the UK housing market that the Bank of England uses to study mortgage policy: large batches of simulation runs finish in a fraction of the time, a validation loss over twenty housing and mortgage indicators scores a run against the data, and a desktop application with a dashboard lets economists run, configure and inspect the model without going near its Java source. [code]

  • Angus Tze Lap Leung (BEng, 2025) built FABS, an open-source C++ simulator for agent-based digital twins of financial markets. It runs a single large simulation in parallel, up to 12.5 times faster than the previous state of the art, and the markets it simulates show the volatility clustering and fat-tailed returns of real ones. The paper appeared at the ACM International Conference on AI in Finance (ICAIF), 2025. [pdf] [code]

  • Quoc Anh Nguyen (MSc, 2025) built Vola-BERT, which turns a pre-trained language model into a forecaster of foreign-exchange volatility by using it as a sequence model instead of a reader, and conditions it on economic events so that the effect of each event can be read off the forecast. It keeps its accuracy with a tenth of the training data. The paper appeared at the ACM International Conference on AI in Finance (ICAIF), 2025. [pdf] [code]

  • Lisa Faloughi (MSc, 2025) built ProtoHedge, a deep hedging agent whose every decision is a weighted average of the actions attached to a small set of learned market prototypes, so that each trade can be traced back to a familiar market scenario, at a cost of under 0.4% in hedging performance. The paper won 🥇 Best Student Paper at the ACM International Conference on AI in Finance (ICAIF), 2025. [pdf] [code]

  • Gene Yu (MSc, 2024) built VCDF, a layer that wraps any causal discovery method for time series and keeps only the causal relations that stay stable across blocks of the data, which makes the underlying method markedly more robust to noise and non-stationarity without changing it. The paper appeared at the Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD), 2026. [pdf] [code]