ABOUT

Built to outsource my own research.

K2 Capital began with one simple goal: to completely outsource my own market research to AI. I wanted to spend twenty minutes a day reading one brief and come away understanding everything that mattered in the markets and theatres I care about — without doing the reading, sorting, and cross-referencing myself. So I built a system that does it for me: agents that ingest the world around the clock, reason about it continuously, and publish their analysis every morning whether anyone is watching or not.

Everything on this site — the daily brief, the dossiers, the live situation room, the intelligence feed — is produced by that network of agents, and I design, build, and operate every layer of it myself. The site is both the product and the proof. Where it goes next: turning the same signal layer into live trading signals — and how, exactly, is the open research question.

Three layers, one autonomous pipeline.

At the base sits the classification layer: more than 300 curated sources are scraped as often as every ten seconds, and every item is classified by an LLM at ingest with a category, a theme from a canonical taxonomy, the affected assets, named actors, and a 1–10 importance score. Classify once, perfectly, and every downstream consumer inherits the quality.

Above it, the synthesis layer: agents active 24/7, continuously re-reading the classified stream and maintaining living situation reports for eleven geopolitical theatres and every major market sector — complete with trajectory calls measured against each morning's baseline, and explicit links between events and the price moves they caused.

At the top, the publication layer: a daily multi-agent pipeline researches, writes, formats, and publishes the morning intelligence brief at 07:00 — end-to-end, no human in the loop. A separate agent rewrites all standing dossiers weekly, archiving every prior week.

Measured, benchmarked, guarded — not vibes.

AI systems fail quietly, so every layer is built to be verified. Classification quality is scored against a hand-labeled evaluation set on every prompt revision, with a frozen multi-thousand-row production snapshot serving as a permanent regression benchmark — no change ships on intuition alone. High-stakes items (low confidence, or importance at the top of the scale) trigger an automatic second-opinion escalation pass before they reach the feed.

The agents also operate under strict anti-fabrication guardrails: structured output validated in code, themes and assets resolved against canonical taxonomies rather than invented, and a standing rule for every writing agent — reporting a price move as unexplained beats inventing a link. No connection is better than a fabricated one.

Published every morning, archived forever.

The Daily Brief is the pipeline's flagship product — a structured read on geopolitics, macro, markets, and crypto, published autonomously every morning since May 2026. The Dossiers are standing intelligence files on the world's live fault lines, rewritten weekly with full historical archives. The Situation Room tracks every theatre live, updated continuously. The Intelligence Globe renders it all on a 3D map — conflict zones, chokepoints, and strategic infrastructure.

Articles are the exception that proves the rule: long-form research written entirely by hand, no agents involved.

Toward agents that reason, grade themselves, and act.

The intelligence stack is the foundation, not the finish line. The next stage of K2 is deepening the agents themselves: self-grading forecasts, where the system scores its own published predictions against what actually happened; richer synthesis, where agents trace causal chains across theatres instead of summarizing them in isolation; and eventually, execution — turning the intelligence into live trading signals, the open question the whole project points toward.

The destination is an operation where the human's role is architecture and judgment, and the system itself — research, writing, publication, and one day positioning — runs on its own. Every layer in between stays documented, benchmarked, and public.