Parallax — Prove the right plan

Parallax architecture: Cloud Run, ADK, Gemini 3.5, Pub/Sub, Firestore, and an evidence-gated executor.

Inspiration

Operational agents often choose one plausible plan before checking the assumptions that could make it fail. In flood relief, a blocked road, unavailable inventory, or an incorrect cost estimate can waste limited time and money.

What it does

Parallax plans a flood-relief operation for ten villages within 72 hours and a ₹1 crore budget. It converts the goal into explicit assumptions, evidence requests, and competing plans. Evidence updates each plan, rejected assumptions can prune a branch, and irreversible actions stay locked behind required proofs and human approval.

The dashboard shows assumption criticality, evidence confidence, provenance, plan selection, action gates, and an audit trail. Custom runs use live Gemini planning. Current probe observations are seeded synthetic inputs and are clearly labeled; they demonstrate orchestration and state transitions, not independent real-world verification.

How we built it

Gemini 3.5 Flash generates structured assumptions, probes, and competing plans.
Google ADK coordinates planning, evidence ingestion, resume, and gated execution tools.
FastAPI runs on Cloud Run.
Pub/Sub handles asynchronous probe events.
Firestore stores persistent run state.
The Next.js dashboard exposes decisions and evidence to the operator.

Innovation

Parallax treats uncertainty as part of execution state. It keeps alternatives active, separates risk importance from evidence strength, maps actions to required proofs, and uses idempotency keys to prevent duplicate execution.

Challenges and learning

The main challenge was separating model-generated planning from execution authority. Gemini proposes structured plans, while deterministic validation and evidence gates decide whether an action is allowed. Reliable retries also required persistent state, event IDs, and idempotent actions.