Things I build to learn

A concept for AI-assisted academic peer review — one platform for publishers, editors and reviewers. It scores every review across six quality dimensions (thoroughness, specificity, constructiveness and more), matches reviewers against OpenAlex's 250M+ works with automatic conflict-of-interest detection, fact-checks citations against their source papers via Semantic Scholar, auto-anonymises submissions, and synthesises multiple reviews into agreements, disagreements and gaps so an editor sees the full picture in seconds.

A concept for an AI-native workspace for cross-border wealth advisers — one adviser running a book of households across the US and India. A team of agents watches each portfolio and surfaces only what needs a decision: a 'Needs attention' queue and an event radar tracking IPO lock-up expiries, RMD deadlines, Roth-conversion windows and ESPP purchases, alongside book-level views of allocation, AUM history and cross-border exposure by jurisdiction.

A 49M-parameter, faithful re-implementation of Moonshot's Kimi K3 — trainable on a single 8GB GPU. It keeps the frontier architecture intact (Kimi Delta + gated MLA attention, a LatentMoE with shared and routed experts, the Muon optimizer) instead of swapping in vanilla Transformer blocks, so you can study a modern MoE model end to end.

Automated ML code generation for tabular data — specialised agents for each pipeline stage, with integrated code execution and debugging.

Core contributor to CortexOn — an open-source agentic framework with tool-calling and orchestration primitives.

Interactive browser visualiser for seven search algorithms — real-time graph construction and heuristics via Python↔JavaScript interop (Brython).

A little directory of interactive physics simulations — explore and play with concepts through a simple browser interface.