I started by proving theorems,Now I build systems that learn.

Showing featured projects. More work lives below the fold.

• Does momentum help or hurt when the objective keeps moving?

  Accelerated PPO Variants · 2026

  Standard PPO uses a clipped surrogate objective to constrain policy updates, but what happens when you add momentum to that? Acceleration mechanisms that work well in supervised learning interact non-trivially with a non-stationary value function and a trust-region constraint. I am implementing and benchmarking accelerated PPO variants against a clean PPO baseline across MuJoCo environments with ten seeds each, tracking not just reward but gradient norms, entropy and policy ratios to understand the optimization dynamics.

  Outcome: Ongoing.

  ResearchEngineering

• Can a policy learn from logged data without ever exploring?

  Offline RL for Autonomous Driving · 2025

  Real exploration in autonomous driving is unsafe; a bad action is not just noisy training, it can mean a crash. We built a two-stage offline RL pipeline in the DonkeyCar simulator: a convolutional autoencoder compressed raw visual observations into a 32-dimensional latent space, then a TQC expert with gSDE generated high-quality trajectories that served as the fixed dataset for a Conservative Q-Learning agent. Ablations on the latent dimension, CQL conservative weight and dataset quality revealed the conditions under which the offline policy succeeds and where it fails.

  Outcome: The expert reached consistent lap completion in ~70k steps. The offline CQL agent recovered stable driving behavior from the fixed dataset in minutes rather than hours of online interaction.

  ResearchEngineeringSafety

  PaperCode

  View diagram

• What does it take to make federated learning trustworthy end to end?

  Federated Learning as a Lifecycle · 2025

  Federated learning is usually treated as a training loop with privacy bolted on. We argued that in regulated settings, healthcare especially, privacy, governance and auditability need to be structural, not ornamental. I co-authored a paper formalizing FL as a four-phase lifecycle: readiness, training, validation and consensus, deployment and maintenance. Each phase transition is machine-checkable through executable policy rules and produces auditable decision records. Validated on multi-label chest X-ray pathology classification.

  Outcome: The framework formalized FL into 4 auditable phases, making every transition traceable through policy checks and decision records rather than informal operational assumptions.

  ResearchSafety

  Paper

  View diagram

• How do you catch and explain pension anomalies without touching the payroll engine?

  Retirement Pension Anomaly Detection and Causal Diagnosis · 2025

  The constraint was strict: no access to the core payroll engine, only its outputs. From 288k records and 214 features, I built an unsupervised detector using a dense autoencoder with per-feature adaptive Huber loss and multi-method thresholding. To move beyond black-box alerts, I added a Bayesian network with a causal hierarchy to attribute root causes and generate traceable explanations auditors could act on. Deployed through FastAPI with a React dashboard and Dockerized services.

  Outcome: 2.5% reconstruction error on payroll data, enabling real-time flagging and root-cause diagnosis while reducing manual verification workload.

  EngineeringSafety

  View diagram

• Can countries build shared flood intelligence without centralizing their data?

  Federated Flood Prediction for Climate Adaptation · 2025

  Hydrological data in Africa is fragmented across countries and institutions, and centralizing it is often politically or technically unrealistic. I contributed to a federated setup where each country trains locally on discharge and station data while contributing to a continental model through aggregation. The system covers flood-year prediction, magnitude regression, and season classification, with both task-specific and shared multi-head architectures.

  Outcome: 2nd place, AI for Sustainable Cities and Communities Challenge, GITEX Africa 2025.

  ResearchEngineeringSafety

  Code

  View diagram

Recognition