Inspiration
Training operations often depend on scattered spreadsheets, attendance records, assessment results, messages, and manual follow-up. The difficult part is not simply identifying that a learner may be struggling; it is deciding what should happen next, coordinating the intervention, tracking whether it happened, and making sure unresolved cases do not disappear.
What it does
CohortPilot is a Taskmaster-style autonomous workflow agent designed to manage that operational loop. It watches incoming cohort data and workflow events, identifies learners or cases that require attention, reasons about the appropriate next action, creates and routes intervention tasks, persists state, and keeps following the workflow until the case is resolved or genuinely needs human escalation.
The target workflow is:
event or new cohort data → detect risk → gather evidence → classify priority → plan intervention → execute or route actions → persist state → monitor completion → retry/recover when needed → escalate exceptions
The goal is to move beyond a dashboard or chatbot. Instead of only telling a trainer what may be wrong, CohortPilot is designed to perform the repetitive coordination work around the intervention process.
Key capabilities
- Ingest cohort, attendance, assessment, and engagement data.
- Detect learner-risk signals and operational exceptions.
- Use Gemini to reason over evidence and select an appropriate intervention path.
- Create structured action plans with priorities, reasons, and next steps.
- Route actions asynchronously so workflows can continue without the trainer guiding every step.
- Persist workflow state and intervention history across runs.
- Track pending, completed, failed, retried, and escalated actions.
- Surface a trainer dashboard focused on decisions that genuinely require human attention.
- Demonstrate failure handling and resumable execution rather than relying on a brittle one-shot script.
How we built it
The project is being built as a modular agentic system rather than a single prompt wrapper. Google ADK provides the agent framework, Gemini 3.5 Flash provides reasoning and decision support, and Google Cloud services provide execution and persistent state.
The planned production architecture separates orchestration, reasoning, action execution, persistence, and user-facing reporting so that failures in one component do not destroy the overall workflow state.
Google Cloud architecture
- Gemini 3.5 Flash — reasoning, evidence synthesis, prioritization, and intervention planning.
- Google Agent Development Kit (ADK) — agent orchestration and tool-driven workflow execution.
- Cloud Run — deployable backend/API and agent service.
- Firestore — persistent learner cases, workflow state, action history, and resumability.
- Pub/Sub — asynchronous action routing and background processing.
- Python / FastAPI — application and integration layer.
Why it is different
Many education AI tools stop at recommendations. CohortPilot is centered on continuous action: detecting a real operational problem, deciding what should happen, executing or routing the work, remembering state, recovering from failure, and keeping the case moving with minimal human hand-holding.
What we learned
The project focuses on the difference between generating a good answer and operating a reliable workflow. Useful autonomy requires explicit state, narrowly scoped tools, observable actions, retry behavior, and clear human-escalation boundaries. It also requires the demo to prove that actions occurred, not merely that the model produced convincing text.
Current build status
This Devpost page is being prepared while the new hackathon build is in progress. Repository, hosted application, architecture upload, testing instructions, and demo-video links will be added after those assets are verified.
Built With
- fastapi
- firestore
- gemini-3.5-flash
- google-adk
- google-cloud-run
- pub/sub
- python
