MetaCode2026
A metacognitive coding benchmark measuring confidence calibration, error detection, and self-correction in frontier models. Submitted to the Kaggle Measuring AGI Challenge.
AI & Machine Learning
A metacognitive coding benchmark measuring confidence calibration, error detection, and self-correction in frontier models. Submitted to the Kaggle Measuring AGI Challenge.
Retrieval-augmented generation pipeline for extracting structured insights from academic papers. Beats naive summarization by grounding responses in retrieved chunks.
ETF classification using K-Means clustering, unsupervised learning applied to portfolio construction. Then layered supervised classification on top to predict category membership.
Ad click fraud detection using XGBoost, imbalanced classification at scale.
Developer Tools
Terminal UI that unifies host processes and Docker container ports in a single view. Eliminates the context-switch between ps aux and docker ps. Idea to shipped in one afternoon.
Real-time cost observability dashboard for LLM API usage, tracks spend, latency, and token burn across sessions. Built because I kept getting surprised by my monthly API bill.
Experiments & Side Bets
Tower-defense loop in Godot 4: elemental totems on concentric rings, waves, seeds and upgrades. Built with Claude Code, Claude Design, and Cursor alongside Godot; playable web soon while the repo stays public.
NFT minting platform for startups, built when on-chain ownership for digital assets felt like an interesting primitive worth exploring.
Linear optimization to select the highest expected-value 11-man squad for the 2018 IPL season.
My first 2D game in the Godot engine. Shipping something playable beats perfecting something unfinished.