Behavioral scientist (PhD, LSE) · I study how people make choices and why they stay committed, using experiments, survey and behavioral data at scale, and AI-assisted measurement checked against human judgment.
📍 London · 🎓 PhD viva 11 Nov 2026 · 🔎 Open to roles in behavioral insights, experimentation, quantitative UX research, public-opinion research and AI-adoption consulting from 13 Dec 2026
| Project | Question | Scale | Key result | Links |
|---|---|---|---|---|
| Branding at the Water's Edge (Paper 3) | How did US news represent each party's position on Ukraine aid? | 37,312 sentences from 11,819 articles, LLM-coded | Partisan gap widened ~40% in three event-linked steps; LLM agreed with human coders at r = .80 | Preprint · Code · Talk at IC2S2 2026 |
| Commenting at the Water's Edge (Paper 4) | Do clearer party signals change how people post? | 217,494 Reddit comments, 5-task coding scheme | Chose a model at ~$0.39 per 1,000 comments (a fifth of the premium option) with 86–89% human agreement; clearer signals made partisanship more visible (~70%) than more intense (~30%) | Data · Code |
| What keeps people committed? (Paper 2) | Which factors predict commitment, and which predict change? | 24,000+ British Election Study respondents | Models using a quarter of 100+ factors more than doubled baseline performance; emotion anchors loyalty, issues and the economy move it | Code |
| Predictors of partisan strength (Paper 1) | What drives party loyalty? | 26,148 records screened across 9 databases | Distilled 50 studies into 56 drivers in 10 themes; led a 9-author team | Published 2024 |
| pi-codex-kit | Can useful features from other coding agents work inside Codex? | 3 hooks + 1 git hook · 2 MCP servers built, 1 integrated · 20 skills (8 adapted, credited) | Brought features from the pi agent and Qoder to Codex as additive add-ons, each with install, verification and rollback steps | Repo |
flowchart LR
A[Question] --> B[Design<br/>experiment or coding scheme]
B --> C[Measure<br/>surveys · logs · LLM coding]
C --> D[Validate<br/>against human judgment]
D --> E[Analyze<br/>ML · causal inference · time series]
E --> F[Communicate<br/>papers · talks · teaching]
- Experiments: factorial online RCTs (e.g. 385 participants on choice design)
- AI at scale: LLM pipelines on 250,000+ texts; models chosen on accuracy and cost
- Open science: public code and data for both text-analysis studies
Teaching (rated 4.8/5 in Social Psychology, school average 4.5; AFHEA) · supervised 5 student research teams, 4 of which won awards · music producer · film photographer