I build software for problems that actually exist.

Working on AWS Shield, which runs on a global fleet of hosts across data centres and edge sites. My project is the automation that finds hosts quietly broken by hardware failure, decides whether the fleet can afford to lose them, and gets them repaired and returned to service without a human driving each step.

• Built automated detection for degraded hosts by correlating hardware telemetry with internal service metrics, so a hardware failure surfaces on its own instead of waiting for a ticket or an incident to reveal it.
• Added a safety analysis in front of every removal — a host only comes out of service once the remaining fleet can absorb its traffic, so mitigating one failure never cascades into an overload of its neighbours.
• Closed the loop end to end — the system raises the ticket to data centre operations, tracks the repair, and returns the host to service once the fault is resolved, turning a manual reactive process into an automated one.

Worked on the enterprise administration platform that internal analysts use to process enrollment. My work centred on an AI-powered enrollment upload feature — the path a multi-family enrollment document takes from being dropped in the browser to being structured data an analyst can trust.

• Cut end-to-end extraction latency by over 95%, from 10–15 minutes down to 30–40 seconds on high-volume multi-family documents, by designing object-oriented Python services around bounded asynchronous concurrent processing with asyncio.
• Built modular backend services integrating Azure Document Intelligence and OpenAI LLM-based extraction into a high-throughput concurrent pipeline, unit-tested with pytest, with per-segment runtime telemetry and partial-failure isolation so one bad page cannot take down a batch.
• Delivered features in an Agile Scrum team using Jira, working directly with project managers and business stakeholders to turn requirements into production-ready systems rather than prototypes.

A student marketplace for the MSU campus. I owned the technical side end to end — architecture, the decisions behind it, and everything that broke at two in the morning.

• Architected the platform in Java and Spring Boot with REST APIs and MySQL behind a service-oriented design, supporting hundreds of concurrent MSU users with JWT authentication and WebSocket-based real-time messaging.
• Led end-to-end software architecture, applying OOP design patterns — Factory and Composite where they genuinely earned their place — along with API orchestration and modular service decomposition.
• Kept the system maintainable enough that debugging and shipping stayed fast as the surface area grew, which mattered more than any individual feature.

Robotic manipulation research. I collected and processed simulation data for UMI gripper tasks in NVIDIA Isaac Sim, turning simulated runs into structured datasets the lab could actually train and evaluate control algorithms against.

• Collected and processed simulation data for UMI gripper manipulation tasks in NVIDIA Isaac Sim, building structured datasets out of simulated environments for training and evaluating control algorithms.
• Analysed gripper performance data and iterated on optimisation strategies to improve manipulation accuracy, refining the data pipelines with the research team and evaluating model output against baseline metrics.

Data and modelling work supporting cross-functional decisions — my first industry role, taken the summer after my first year.

• Built data analysis pipelines with PyTorch and scikit-learn to preprocess large datasets and train clustering and classification models, delivering insights that fed directly into cross-functional business decisions.
• Developed interactive dashboards and structured reports that translated complex model output into clear visualisations for stakeholders, which is the half of ML work that usually gets skipped.