Now

Between roles. I left Allianz Global Investors on 28 February 2026 after four years architecting AI systems and data pipelines across a €176B Fixed Income platform. The last push before leaving was the firm's first enterprise agentic AI system, shipped to user testing within the COO's December deadline roughly two months after build start.

Job searching for applied AI, quant, and data roles. Paris primary, London close second; open to sponsorship. I'm mainly looking at applied AI, AI/ML engineering, quantitative development, data science, and forward-deployed engineer paths, where system architecture meets products people actually use. Floor is £65K / €60K with flexibility for equity and benefits. I'm running my own search through JobHunter, which scores 500 to 1,000 roles a day.

Hardening Lineup's data platform. Lineup is in production: 1M+ events, 300K+ artists, 122 unit tests, zero data losses in twelve months. Current work is refactoring the pipeline into an installable Python package with proper dependency management, adding incremental loads to cut BigQuery costs, and swapping the 57MB JSON cache out for Redis. Next up: Looker Studio dashboards for pipeline health and attribution modelling for venue popularity.

Closing out NutriPlan v1. FastAPI backend is complete: meal generation, swapping, feedback, and household coordination. Calorie accuracy within ±10%, generation under two seconds end-to-end, 99.8% uptime on Render. SwiftUI frontend is design-complete and implementation is underway; targeting a TestFlight beta by summer 2026. After that, Spoonacular/Edamam for real recipe data and a freemium ladder: free calorie tracker, premium meal planning, pro macro/household tuning.

Earnings Call Analyser research. Started Q1 2026. Validating Chiang et al. 2025 on whether Q&A alignment in earnings calls predicts forward equity returns. Architecture is done: multi-source transcript ingestion, FinBERT (768-dim) embeddings, PyTorch contrastive classifier with a combined loss (cross-entropy + MSE + SimCSE-inspired), FastAPI backend, Streamlit dashboard, Kubernetes deployment. Near-term blocker is labelled training data: 100 to 200 manually scored Q&A pairs to train the classifier before backtesting against 2020 to 2025.