Danny Chait

Artifacts

Whiteboards, diagrams
and the sketches behind them

Architecture, timing models, dashboards, and the paper a feature started on. Click any of them to open the full-size original.

License system architecture diagram: CRM and Navision synchronise with the License Editor, which issues licenses and license types to the BBP app. The BBP app resolves those into enabled session types on the B1 hardware unit.
How a commercial record in CRM becomes an enabled session type on the hardware.
B1 License System
TrackMan B1 usage dashboard with filters for date range, session type, organization, app version and firmware, showing 1,075 total units, 146K sessions, 7.968M pitches tracked and 3.108M hits tracked across 552 organizations.
The fleet-wide view used to spot dormant units and target renewals.
B1 License System
TrackMan B1 iPad app during a pitching session, with pitch-by-pitch metrics on the left, release and spin numbers in the centre, and synchronised Edgertronic camera footage playing on the right.
The end state: pitch metrics and synchronised high-speed video in a single session view.
Edgertronic Integration
Sequence diagram with lanes for Pitcher, Tracking Unit, iPad and Camera. The pitch is thrown and analysed, a launch packet with release timestamp reaches the iPad within 500ms, the camera receives it 1.5 to 2.5 seconds later, triggers, selectively saves the frames around the release frame, stores the video with a tracking ID, and returns it to the iPad.
The latency budget between release, trigger and frame selection is what makes or breaks capture rate.
Edgertronic Integration
Four-stage architecture. Sources: raw MongoDB collections covering practice sessions and shot events, completed rounds, AI assistant requests, in-app feedback and device registrations, with internal accounts stripped on read. Rollup: a MongoDB Atlas scheduled trigger produces daily usage totals split by club and by app version, plus a matching rollup for the second product line, with day boundaries fixed to one timezone, per-metric coverage floors, blanks instead of zeros, and idempotent backfill. Single source: a handful of Firestore documents, gated by security rules. Serve: a Cloudflare Pages dashboard site and a Cloudflare Worker MCP server, both behind a Google SSO gate, reaching colleagues in a browser and AI agents in conversation.
The platform end to end. Everything converges on a handful of documents before it fans back out.
Usage Platform
Left: the tool list as a menu. Orient, with list_datasets. Composed tools that answer a product question outright: get_product_briefing, find_movers, get_feature_adoption, get_activation_funnel. Primitives beneath them: get_kpis, get_timeseries, get_breakdown, compare_periods, get_metric_catalog, get_definitions. Right: the design premise, that an agent will misread this data confidently unless stopped, and six guardrails enforced in the tool layer: distinct users do not sum, a blank is not a zero, a missing row is a zero, a rate is not a percentage, unequal windows are withheld, and no funnel is faked.
The tool list is the interface. The client panel shows names only, so it reads top to bottom as a menu of questions.
Usage MCP Server
A domain-restricted Google sign-in leads into a dashed boundary marked "inside the browser, nothing below leaves the device", containing four steps: report header with name, department, date and business purpose; line items across airfare, hotel, dining, transport and other; receipts held as local blob URLs with no upload; and assembly with jsPDF, producing a header and itemised table page followed by paginated receipt pages. A note explains that every receipt is redrawn through a canvas first, because a production content security policy can block fetch and FileReader on blob URLs, and re-encoding guarantees the declared image format matches the actual bytes. Output: a finished PDF named for the employee and report date.
The whole tool. The dashed boundary marks what never gets uploaded.
Expense Generator