Inspiration
AI filmmaking has made asset generation fast, but releasing a coherent film is still an operations problem. Scripts, character references, keyframes, voice, clips, captions, continuity checks, and the final export often live in disconnected tools. A team can generate impressive shots and still be unable to answer a basic question: is this exact cut complete, consistent, and safe to premiere?
Backlot treats that question as the product.
What it does
Backlot turns one greenlit premise into a persisted short-film production. A specialized AI crew handles writing, script review, production design, casting, storyboarding, audio, keyframes, clips, continuity review, editing, and release.
The Studio Floor shows each department's state, generated artifacts, bounded agent handoffs, continuity findings, render budget, and recovery actions. The final film enters the public Theater only after the Release Observer verifies production-specific evidence through the official Grafana MCP server.
The release gate checks that:
- every required shot was rendered;
- continuity quality control passed;
- the export uses the approved media lineage; and
- the final edit exists and passed editor checks.
If telemetry is missing, a provider fails, media becomes stale, or a readiness check is false, Backlot holds the premiere and explains why. It never replaces a failed production with synthetic success.
How we built it
Backlot combines agentic reasoning with deterministic workflow control:
- Google ADK and Gemini power specialist agents and typed handoffs.
- A deterministic ProductionOrchestrator owns state transitions, retries, budgets, file writes, authorization, and publication.
- Vertex AI / Gemini generate structured production artifacts and media inputs.
- FFmpeg assembles audio, captions, clips, credits, and the final film.
- Prometheus and Grafana expose production health and release-readiness signals.
- The official Grafana MCP server lets the Release Observer discover the production dashboard and query evidence for the exact production before release.
- FastAPI provides the production API, while Next.js, React, and Tailwind CSS power Greenlight, Studio Floor, and Theater.
Every accepted agent handoff records the production, responsible agents, typed decision, artifact hashes, tool metadata, recipient, and timestamp. Backlot stores decisions and provenance—not private chain-of-thought.
The media lineage is explicit: keyframe revision and hash → rendered clip revision and hash → encoded-frame continuity result → final export membership.
Changing a keyframe invalidates its dependent clip. A continuity failure triggers a bounded reshoot, and the editor rejects stale or unverified media.
Why Grafana is load-bearing
Grafana is not a dashboard added after the build. It participates directly in the product decision.
Immediately before publication, the Release Observer uses Grafana MCP to find Backlot's production-control dashboard and query four Prometheus readiness series scoped to the current production ID. The evidence—including dashboard, datasource, query, named results, scope, and decision—is persisted. If any required result is absent or false, Theater publication is blocked.
This turns observability from passive monitoring into an active release contract.
Challenges we ran into
Keeping continuity attached to shipped bytes
An early pipeline could regenerate a failed keyframe without proving that the rendered clip and final edit changed with it. We solved this by binding keyframe, clip, sampled-frame QC, and export membership through revision hashes.
Combining agents with durable execution
Giving an LLM authority over state, spend, retries, and publication made recovery fragile. We kept agents responsible for creative and specialist decisions while a deterministic orchestrator controls every irreversible transition.
Preventing duplicate paid work
Greenlight, resume, cancellation, recovery, and sequel requests share an idempotent admission boundary with capacity checks, persisted attempts, and bounded retries.
Accomplishments that we're proud of
- A complete greenlight-to-premiere product loop rather than a prompt-box demo.
- A fail-closed release gate driven by real Grafana MCP queries.
- Typed, auditable multi-agent handoffs without exposing chain-of-thought.
- Artifact-level continuity and provenance that survive retries and reshoots.
- Clear failure states: an unhealthy production stays private and tells the operator what must be fixed.
What we learned
Agent delegation becomes reliable when outputs are typed, persisted, and accepted by a separate authority. Media continuity is not only a visual-scoring problem; it is a lineage problem. Most importantly, observability becomes product infrastructure when a real user action depends on its evidence.
What's next for Backlot
Next, Backlot will validate the workflow with small creative teams, strengthen distributed production execution, add richer reviewer approvals, and measure whether evidence-backed release review reduces failed exports and coordination time. The goal is not to become another general video generator—it is to become the release-operations layer for AI-assisted media.
Log in or sign up for Devpost to join the conversation.