dreamatlas.studio / projects / event-producerCasefile Mounted · 004

004 / Event Producer

Event Producer

Agentic casefile system

Event Producer is an agentic event-production casefile system built as a Kaggle x Google AI Agents capstone for the Agents for Business track.

It turns event basics and a brief into a saved, costed, scheduled, approval-gated working casefile. A brief can drift, but the casefile has to stay true.

Agents propose and draft; deterministic engines compute; humans approve consequential action.

Watch the capstone walkthrough, then use the tabs to inspect the state model, agent crew, deterministic engines, and safety boundaries behind it.

// SURFACES - casefile records

The working room and the agent crew.

Browse the public casefile surface, then switch records to inspect the specialist crew behind it.

// select record

AI Producer / Orchestrator

Routes broad requests over the saved casefile and returns typed proposals.

Requirements Agent

Extracts useful context and reconciles it against structured event facts.

Creative Concept Agent

Drafts event directions, titles, experience beats, and additions.

Scope Strategy Agent

Suggests additions, cuts, reductions, and tiering tradeoffs.

Vendor Copy Agent

Drafts editable vendor copy from the selected vendor context only.

Risk Review Agent

Flags budget pressure, schedule risk, vendor chase items, and operational gaps.

Budget Engine

Computes spend, contingency, headroom, and scope rollups deterministically.

Approval Wall

Blocks gated external or state-changing actions without human approval.

// ARCHITECTURE - bounded agentic workflow

A casefile, not a chatbot.

Event Producer captures structured event facts, saves them as casefile truth, reconciles free-text briefs against those fields, and lets specialist agents generate bounded work products.

Browser / Paper War RoomCasefile selector, event basics, brief intake, requirements, Mission Control, scope, budget, run sheet, Vendor Notebook, approvals, audit.

FastAPI backendTyped contracts, casefile store seam, specialist agents, deterministic engines, vendor notebook, structural approval gate.

Storage modesLocal JSON for clones and Firestore for the hosted demo environment.

Agent boundaryAgents handle language, proposals, risk framing, and drafts. They do not own arithmetic, schedule mechanics, sending, or payments.

Engine boundaryBudget and run-sheet logic are computed by deterministic code, not free-text LLM math.

Demo boundaryPublic casefile isolation is browser-session-based and not production authentication.

// BUILD JOURNEY - from capstone attempt to trust-first casefile

A convincing AI dashboard is useless if it plans the wrong event.

The trust failure is specific: a brief may say 50 people, the producer updates the casefile to 100, and generated outputs still behave as if 50 is true.

// select record

Event Producer began as a capstone attempt at agentic event production, then reset around one rule: a strict precedence order for what counts as true.

User-entered structured fields

User-edited saved casefile values

AI-extracted brief values only when blank

Explicit missing or unknown state

Seed demo defaults only inside seeded demos

State truth and local casefilesFile-backed casefile store, index, timeline, and artifacts, with resolved-state precedence and regression tests.

Direct specialist agentsCreative Concept, Scope Strategy, Vendor Copy, and Risk Review actions over the saved casefile.

Mission ControlA runtime-loaded skill-card registry driving role cards, direct actions, and contract tests.

Vendor NotebookA persistent per-vendor chase list with editable drafts, logs, statuses, and injection-screened replies.

Post-polish correctionsCurrency truth, smarter scope quantities, sane run-sheet timing, and clearer budget behavior.

The harness matters more than raw model output.

Agents should not own arithmetic or schedule invariants.

Human approval should be structural, not prompt-based.

Event Producer is not an autonomous planner. It is an agent-assisted production workspace — specialist agents, deterministic engines, and human approval working together without sacrificing control.

// EVIDENCE - running build log

What shipped, and when.

Appended from the project changelog as milestones land.

2026-07-07SUPPORTPost-submission demo hardening — Firestore error handling, Firebase Hosting fixes, and larger model-output budgets tightened for the public hosted instance (after the Kaggle deadline).

2026-07-06RELEASECapstone submitted — final sprint closed out honest budget metrics, seeded demo casefiles, and UI polish ahead of the Kaggle deadline.

2026-07-05FEATUREVendor Notebook and Mission Control shipped — persistent per-vendor chase lists with injection-screened replies, and a registry-driven Production Crew board.

2026-07-01SUPPORTCurrency and schedule-truth pass — casefile currency threads end-to-end through pricing and agent prose; run-of-show anchors to the real event date.

2026-06-30SUPPORTSkill cards made load-bearing — every agent's live prompt assembles from the same versioned role-card registry the API serves, pinned by contract tests.

2026-06-22PLATFORMDeployed to Cloud Run and Firebase Hosting — FastAPI backend, static Next.js frontend, and CI QA gates ahead of every deploy.

2026-06-22FEATUREAgent crew and orchestrator wired — five role agents plus a top-level orchestrator routing requests over the saved casefile.

2026-06-21FEATUREDeterministic engines and state-truth model — a Decimal-only Budget Engine, a CPM Scheduler, and a local casefile store with source precedence, so agents draft but never own arithmetic or schedule truth.