Work

Research and prototyping of agentic AI systems; evaluation of frameworks and approaches.

• Designed a Production-Informed Accelerated Multi-Agentic Simulation System: production-derived failure distributions, probabilistic environment and dependency simulation, agent-trajectory and multi-stage workflow evaluation, harness validation and recovery testing, low-cost accelerated execution using smaller models and cached dependencies, anomaly discovery for behaviours absent from historical datasets, and release gates with automated rollback policies.
• Established the feedback loop that makes the above self-correcting: production execution → telemetry and trajectories → streaming analysis → learned failure distributions → stochastic simulation experiments → evaluation gates → safer release → new production evidence.
• Designed and built an LLM observability and reliability harness to make non-deterministic systems safe in production. Improved structured-output success from ~71% to near-100% through prompt hardening, validation layers, normalisation and controlled recovery — a deterministic contract enforced over a non-deterministic model.
• Explored end-to-end observability across LLM pipelines (logs, metrics, traces) to detect schema drift, malformed output, latency variance and downstream failure, and investigated control-loop patterns — observe, detect, intervene, stabilise, measure — with failure-aware recovery: classification-based retries, bounded recovery, and cost/latency trade-off management.

Real-time, safety-critical, physical-AI machine learning.

• Built a machine-learning platform that scaled brain–computer-interface model training and evaluation 100×, to billions of neural data points.
• Created BCI software and an ML model for brain-to-text communication; the system was presented at the Neuralink 2022 Show and Tell.
• Delivered real-time, partially observable signal pipelines under hard latency and safety constraints — directly analogous to physical-AI and edge-agent control loops (sim-to-real, graceful degradation).

Real-time scoring at scale; adversarial defence; internet-scale reliability.

• Designed and implemented automated anomaly detection in near real-time streaming and ML scoring at a scale of billions of events per day, with peaks of 50K QPS.
• Led the research, design and implementation of a bot-detection system countering adversarial attacks that affected more than 40% of Sift's traffic — adversarial-agent defence.
• Designed and implemented a Rapid ML Modeling system that improved modelling and prototyping speed by more than 10×.

Large-scale distributed machine-learning platform engineering.

• Designed and built a distributed ML platform supporting billions of events and hundreds of thousands of features, with automated feature extraction and hyperparameter search; reduced existing model-training runtime by more than 10×.

Real-time video understanding and applied deep learning.

• Built a real-time video-processing mobile app with analysis and classification by deep neural networks; the app automatically generated videos that were accepted and nominated at international film and fashion festivals in Denmark and New York.
• Led R&D on social-media and news sentiment correlation with the S&P 500, and a deep-learning platform matching startups with angel and VC funds.

Regulated, high-consequence financial infrastructure — irreversible actions, financial-grade correctness.

• R&D of UBS mortgage, loan and fixed-income derivatives, Risk/PnL and trading platforms — a front-to-back solution for fixed income and mortgage-backed securities businesses, with total combined notional on the CDS contracts in the range of $10T+.
• R&D of back-office platforms including subledger and clearing & settlement; led 50+ FTE and $10M budgets.