Agentic Cinema Hackathon · ClickHouse Track
Prod_Mode: Active
StudioDirector
BrightKin Studio Director turns production data into auditable decisions. Its new Studio Mesh layer lets the Director coordinate Compliance, Greenlight, and Release specialists over one ClickHouse production memory. Demo data verified 2026-08-01 UTC.
- interwoven agents
- 4interwoven agents
- queried warehouse
- 1queried warehouse
- synthetic snapshot
- 1synthetic snapshot
DIRECTOR ⇄ COMPLIANCE ⇄ GREENLIGHT ⇄ RELEASE ⇄ CLICKHOUSE
ISO 800NET STATUS: OPTIMALHow it works
One question in. A routed, grounded answer out.
- 01 · input
Ask
Ask a plain-language question about an episode or a track. No SQL, no dashboard hunting, no knowing which table holds the answer.
natural language in
- 02 · logic · active
Studio Mesh orchestrates
The Director coordinates Compliance, Greenlight, and Release specialists. Each owns a decision boundary; routing and evidence stay visible.
agent handoff
- 03 · output
Grounded answer
Every answer is routed through a ClickHouse query against a clearly labeled synthetic snapshot. The UI exposes the query used for grounding.
query-grounded
Current session timecode 00:00:00:00
Try it
Put a question to the Director.
This runs the deployed agent and a fresh ClickHouse query against the bundled synthetic snapshot (verified 2026-08-01 UTC). Each answer shows which agent handled it and the query used for grounding.
studio-director · session
Pick an example question, or ask your own. The Director decides whether it can answer - or whether Compliance should.
Beyond the demo
Beyond the Demo — The Agent Pattern
This demo answers three fixed questions about one fictional slate. The mechanism underneath it isn't specific to animation: a routing agent that answers directly or delegates to a specialist sub-agent, both grounded in parameterized ClickHouse queries against a live warehouse, with the exact SQL returned alongside the answer. Anything that already logs stage/status events and structured attributes somewhere queryable can run this same loop.
VFX/render QC copilot
Same stage/status schema (item_id, stage, status, ts, notes), aimed at a render farm instead of a script pipeline. “Why is shot_042 blocked” becomes a Director-style lookup; a Compliance-style sub-agent checks delivery specs — color space, frame rate, codec — before a shot ships, the same pure-function pattern as check_pacing.
Rights-clearance sub-agent
check_music_policy already reduces to one question: does this track's provenance attribute say ‘original’? Point the same check at a real rights-clearance status field and a music supervisor gets a pass/fail with the exact row it read, not a paraphrase.
Multi-vendor delivery gate
item_id/stage/status generalizes to shots handed to outside vendors. The Director answers ‘what's blocking shot X at Vendor B,’ and a Compliance sub-agent enforces the studio's delivery standard — frame rate, aspect ratio — per vendor, per shot.
Slate-wide representation audit
check_diversity today runs against one item_id at a time. list_latest_items already aggregates every item's current state for the dashboard — wiring that into the compliance path turns ‘is episode_1 compliant’ into ‘which items in the whole slate aren't,’ in one pass.
Trailer-cut compliance check
Before a marketing team locks a promo cut, the same check_pacing / check_music_policy functions run against the exact clip set they selected — no new logic, just a different item_id list handed to code that's already unit-tested.
Per-studio compliance SaaS
The tenant_id column and the ‘bring your own show’ intake path already scope every read and write to one browser-generated id. Swap the client-generated UUID for real authentication and the same isolation model becomes a real multi-studio product, not a demo convenience.
Now vs. next
What's actually built, and what isn't — honestly.
| Area | Now | Next |
|---|---|---|
| Compliance scope | One item_id checked per question | Slate-wide sweep across every item in one pass |
| Agent writes | Read-only — agents report findings, never log them | Agent-initiated INSERT (e.g. auto-flag “pacing_review_needed”) |
| Tenant isolation | Client-generated UUID in localStorage, no login | Real authentication gating each studio's rows |
| Data freshness | Request/response — query runs once per question | Live view that re-queries as new events land |
| Standards definition | Three rules hardcoded as Python functions | Studio-configurable rules stored as data, not code |
This is just the beginning
See the code. Fork it. Make it yours.
Every agent, tool call, and parameterized query described above is in the public repo — two ADK agents, the ClickHouse tool layer, and the compliance functions, all real, all readable, no placeholder boxes.