things i've done (extended version):
open source TTS @ MetaVoice
- petabyte-scale data pipelines, owning infra and data
- core model architecture experiments
- task evaluation pipelines
identity verification @ Onfido
- VLM LoRA in production @ scale in 2023 (with talk @ London AI Tinkerers)
- used 97.5pp less resources with a performance regression of < 1pp. Quickly unblocked a critical delivery (due to GPU scarcity) & saved a projected $2M p.a
- fraud detection
- shipped flagship general visual fraud anomaly detection that reduced FPR by ~20pp @ 1% FNR & contributed to the visual fraud risk engine to optimise at an engine-level, increasing robustness.
- document extraction patent 11657631
- shipped flagship template-based OCR extraction system that increased automated coverage by over 600% @ improved accuracy. Co-authored associated tooling to validate performance & optimise system-level parameters which increased onboarding bandwidth by 10x.
- VLM LoRA in production @ scale in 2023 (with talk @ London AI Tinkerers)
an experiment in education - authored & taught a 3 month compsci course
founder @ Aparien
- created a 360 review HR SaaS platform with an emphasis on modern UX to produce anonymised, weighted and actionable results per user with aggregated analytics for management. Used by a handful of consultancies for their respective clients who paid per seat for platform usage.
SWE on amazon air EU