Multi-agent systems · October 2026
MatchMind
Built for the Microsoft Premier League Hackathon 2026. Five agents turn football questions into checked claims over a simulated match event log, with explanations tailored to beginners, fans, and analysts.
A deterministic Skeptic recomputes claims from events. Unsupported numbers are rejected, and model failures fall back to rules. Built with Microsoft Agent Framework, FastAPI, Next.js, and Azure Container Apps.
View repository →Wireless research · Ongoing
Lightweight neural channel estimation
Investigating compact neural estimators for OFDM under noisy conditions and limited compute. Built reproducible simulations, classical baselines, and repeated-seed training with Sionna and PyTorch.
In the documented three-seed study across 15 channel and noise settings, the neural model reduced normalized mean squared error by about 79–87% versus least-squares linear interpolation. The covariance-informed classical estimator remained strongest. These are bounded experimental results, not a general performance claim.
View repository →Network diagnostics · 2026
NetLens
A local-first assistant for unreliable internet. Measures latency, packet loss, and DNS timing, then scores connection health and readiness for tasks such as calls and uploads.
Python, FastAPI, SQLite, and JavaScript. A rule engine provides practical recommendations, while saved scans make connection history visible.
View repository →Research direction · Ongoing
Agentic inference profiler
Exploring how to measure the cost of multi-step agents: model and tool calls, latency, context growth, and inference cost.
An early research direction focused on identifying bottlenecks and informing model selection, prompting, and system design.
View repository →Machine learning · Course project
Solar power generation forecasting
Compared regression models for 15-minute-ahead solar generation forecasts. Used chronological evaluation and exact timestamp matching to avoid mixing past and future observations.
Python, Pandas, NumPy, scikit-learn, and Jupyter. The study documents timestamp gaps, incomplete inverter reporting, and the limits of a roughly 34-day dataset.
View repository →