The desk closes the day with a video.

Two hours after the closing bell, a workflow wakes up, pulls the session's real figures, builds a deck, generates fresh AI-presented clips speaking that day's numbers, assembles a narrated recap, archives it, and publishes it here. No human edits the clip. The video below is the most recent session; the archive underneath is every session since the desk started.

Paper account. The desk runs on a simulated (paper) trading book. This is an engineering demonstration of an agentic pipeline, not investment advice and not a live-money track record.

Numbers first, and they must reconcile

The pipeline starts from data, not narrative. A post-close extraction reads the day's equity, P&L, fills, open book, and the decision ledger — every trade tied to the signal that justified it — and the headline P&L must reconcile against the account's own close-to-close equity change before anything renders. A recap with wrong numbers is worse than no recap, so a mismatch stops the run.

Fresh clips, today's words — enforced, not requested

The presenter segments are generated new every session: an agent on a render node drives the video tool itself, animating a headshot with a script that speaks that day's actual figures. Yesterday's clips are deleted before the build starts, so reusing them is structurally impossible — if fresh clips don't exist at collection time, the run fails instead of shipping something stale. The deck scene is narrated by the video tool's own voiceover from the same data.

Assemble, verify, publish

The host stitches the sting, the presenter clips, and the narrated deck into one video, renders the deck to a PDF report, emails the finished cut, appends the dated artifacts to a git-versioned archive, and deploys the site — the same artifacts you see on this page. Weekends and market holidays are skipped by checking the exchange calendar: no session, no recap.

The wider story — signals, risk gates, and the ledger behind these numbers — is in how the trading swarm justifies a trade, and the live board is on the trading page.

Every session, archived

Each entry links the video, the deck it was built from, and the nightly PDF report.

On the framework: the recap is a scheduled ticket workflow on the oshal runtime — connector-brokered market data in, a bot-owned store for positions and reasons, the render dispatched to a swarm node, and every model call in the central cost ledger.