Selected work across data, policy, AI, and tooling. Each project focuses on a real-world problem — misinformation, worker transition, diagnostic safety, climate resilience, or SME operations — and tries to build something practical.
Explore arms, mechanisms, and KPIs of the radiology AI assistance RCT.
Problem: It’s very hard to study coordinated behaviour and narrative shaping on X in Persian/Farsi, especially when accounts mix bot-like activity with human signalling.
What I built: I’m building a Chrome extension that scrolls a profile, extracts posting patterns, detects Farsi/RTL content, and surfaces red flags like abnormal timing, repeated messaging, or engineered engagement.
Problem: Clinicians under time pressure make life-impacting diagnostic calls. The question is: does AI actually improve decision quality, or just add noise and overconfidence?
What I built: I designed an experimental protocol (multi-reader multi-case) to test whether AI suggestions and AI confidence scores change clinician accuracy, calibration, and diagnostic bias in pneumonia detection.
Problem: Micromobility adoption in London is uneven across income and geography; TfL needs reliable measures and practical levers to improve equitable access.
What I built: Built a measurement framework (exposure, usage, accessibility), linked area indicators to trip data, and designed behaviorally informed interventions with an equity screen.
Problem: Late-night digital media use leads to widespread bedtime procrastination, harming sleep quality and productivity among students.
What I built: We ran a randomized trial at LSE testing Mental Contrasting with Implementation Intentions (MCII). The intervention reduced bedtime procrastination by 0.6 points, cut pre-sleep media by 30 minutes, and improved satisfaction with bedtime control by +1.34 points.
Problem: Climate policy in Eastern Europe and Central Asia often talks about ‘resilience,’ but doesn’t always specify social protection for workers, gender risks, or just transition mechanisms.
What I built: I built a live tracker that maps policies, highlights gaps (gender, disaster risk, vulnerable groups), and flags where social protection is missing from climate commitments.
Problem: For SMEs importing tiles/sanitary ware, sales depends on WhatsApp/Telegram conversations. Reps waste time sending the same specs, photos, and prices again and again.
What I built: I built a Telegram bot that serves as a searchable catalogue. It returns product cards (image, specs, sizing, finish) instantly from a CSV instead of manual replies.
Problem: AI is reshaping labour demand, skill requirements, and job quality, but policymakers mostly see headlines, not structured evidence.
What I built: I work on skill-level and task-level evidence: how AI exposure affects employability, wage pressure, and transition risk — especially for workers in lower-signal labour markets.