Built in the open.

StefanOjanen/worksidian

Worksidian

Parallel Claude Code sessions with a shared brain.

Shell · MIT

Run several Claude Code sessions against one codebase without them colliding: git worktrees isolate the edits, a shared status file lets the sessions find each other, and an Obsidian vault keeps what they learn.

• ·One coordinator in the main checkout, plus axis workers scoped to modes of work, bootstrapped by a SessionStart hook
• ·A gitignored team-status file as the coordination substrate: on disk, editable by hand, survives every restart
• ·Decisions, plans and research accumulate in the vault and load into every session through a generated context map

StefanOjanen/hubs-for-free

Hubs for Free

Interpretability research on what attention heads share.

Python · MIT

The standard head-interaction analysis finds a coordinator head and a spectral gap even in random matrices. In trained models it finds them for a different reason: a layer's heads are, to first order, one shared operator, and that operator is the attention sink. The repo follows the thread from a null result to a one-parameter law that holds from 0.1B to 8B parameters across ten architecture families.

• ·Two short proofs explain why untrained attention already shows a hub and an eigengap
• ·Across 230 high-sink layers of eleven models, the shared component is the ideal sink operator
• ·Preregistered tests, and the instruments released as a toolkit anyone analysing attention can run

StefanOjanen/catalyst-mosaic-investing

Catalyst-Mosaic Investing

A multi-agent research desk for a concentrated portfolio.

TypeScript · MIT

A Claude Opus portfolio manager orchestrates parallel Sonnet research workers with hard output contracts, synthesising alt-data, filings and options flow into catalyst setups whose exit rules are committed before the event.

• ·Four to five research workers per daily run, each owning one domain, plus a deep-dive worker spawned only when needed
• ·Exit rules locked at entry and executed mechanically, with a calibration log of predicted against actual
• ·A local dashboard: live-marked book, risk and exposure views, and a catalyst calendar

protocolengine.io

ProtocolEngine

Health protocols graded by the studies behind them.

Founder · 2026

An evidence-led health intelligence platform that turns published research into personal health protocols. Built solo: habits, foods and supplements are graded against the research, and every grade traces back to specific papers.

• ·Daily ingestion from PubMed, bioRxiv and medRxiv: 43,000+ papers processed into 230,000+ typed evidence claims
• ·Eleven specialised AI agents in production, and a citation-grounded RAG Evidence Assistant
• ·Next.js, TypeScript, Supabase, pgvector and Anthropic Claude