FlowPilot: AI Agent for Engineering Team Productivity

Proactive autonomous AI agents boosting engineering team productivity and developer velocity by detecting PR bottlenecks, unblocking sprints, and reclaiming focus hours with Google Gemini.

Comment

Inspiration

Software engineering teams today lose thousands of productive hours to meeting overhead, context switching, and silent PR blockers. Engineering managers and tech leads spend up to 30% of their week manually triaging crowded calendars, chasing status updates in Slack, deciphering GitHub alerts, and unblocking stalled sprints.

Traditional engineering management tools are purely passive dashboards: they tell you a sprint is behind after the deadline has already passed.

We were inspired to build FlowPilot to transform engineering operations from reactive fire-fighting into proactive, autonomous intelligence. We asked ourselves: What if an intelligent AI agent could continuously monitor engineering telemetry across GitHub, Jira, Slack, and Google Calendar, detect bottlenecks before they derail sprints, and autonomously orchestrate precision calendar and workflow interventions to protect developer focus time?

What it does

FlowPilot is an autonomous multi-agent AI system that proactively boosts engineering team productivity by unblocking PRs, eliminating meeting bloat, and recovering developer focus hours.

Key capabilities include:

  • Multimodal Engineering Telemetry Ingestion: Continuously captures and normalizes signals from GitHub PRs, Jira sprint backlogs, developer Slack check-ins, and Google Calendar schedules.
  • Coordination Risk & PR Lag Scoring: Algorithmic scoring engine that calculates real-time Coordination Risk Scores (0–100) and PR review wait times to identify team vulnerability before sprint deadlines.
  • Autonomous Calendar Interventions (Google Calendar API): Intelligently evaluates scheduled team ceremonies and converts low-value meetings into Asynchronous Standups, dedicated Deep Focus Blocks, or Emergency Swarms, directly recovering wasted hours.
  • Human-in-the-Loop (HITL) Governance: Empowers engineering managers with a 1-click "Approve & Execute" console featuring transparent root-cause rationale and estimated hours saved.
  • Policy Guardian Safety Gate: A deterministic policy enforcement layer guaranteeing the AI never alters client demos, 1:1s, or executive meetings, eliminating hallucination risks.
  • Tamper-Proof Audit Inspector (Google Cloud Firestore): Logs every signal, prompt hash, safety check, and calendar modification for complete enterprise compliance.
  • Gemini-Powered Monthly Engineering Retrospectives: Automatically synthesizes 30-day velocity patterns, meeting-to-code ratios, and recurring bottlenecks into executive-ready reports.

How we built it

FlowPilot is built around a specialized 8-agent autonomous architecture powered by Google Gemini and Google Cloud Platform:

The 8-Agent System:

  1. Ingestion Agent: Normalizes raw engineering signals across GitHub, Jira, Calendar, and Slack.
  2. Evidence Agent: Grounding layer attaching verifiable URLs, commit SHAs, and ISO timestamps.
  3. Health Agent: Computes quantifiable Coordination Risk and PR Lag metrics.
  4. Cadence Agent: Invokes Gemini 3.5 Flash via the Google GenAI SDK to synthesize optimal interventions.
  5. Policy Guardian: Strict safety gate verifying proposed actions against deterministic rules.
  6. Executor Agent: Dispatches real-time, idempotent updates via the Google Calendar API.
  7. Retro Agent: Synthesizes 30-day cross-sprint organizational patterns using Gemini.
  8. Outcome Agent: Tracks cycle time velocity gains, unblocked PRs, and hours saved.

Tech Stack & Cloud Infrastructure:

  • AI & Reasoning: Google Gemini 3.5 Flash (gemini-3.5-flash) via the official google-genai SDK.
  • Cloud Infrastructure: Google Cloud Run for serverless container execution and Google Cloud Firestore for state/audit persistence.
  • Security & APIs: Google Cloud Secret Manager for zero-hardcoding credential hygiene and Google Calendar API (v3) for live scheduling updates.
  • Full-Stack Application: FastAPI (Python 3.11) backend + React 18 / Vite / Tailwind CSS frontend bundled into a multi-stage Docker container.

Challenges we ran into

  1. Eliminating LLM Hallucinations in Calendar Modifications:
    • Giving an LLM open-ended write permissions to enterprise calendars is risky. We solved this by architecting the Policy Guardian layer between the Cadence Agent (reasoning) and the Executor Agent (action), ensuring deterministic policy validation with zero hallucination.
  2. Domain-Wide Delegation & Google Calendar API Sync:
    • Setting up Service Account authentication to seamlessly patch recurring corporate calendar events without OAuth consent fatigue required careful implementation of service account delegation and Secret Manager integration.
  3. Cloud Run Multi-Stage Docker Packaging:
    • Compiling a lightweight single container serving both the Vite React SPA frontend and the asynchronous FastAPI backend while preserving hot routing and Cloud Run port standards (PORT=8080).

Accomplishments that we're proud of

  • Sub-Second Multi-Agent Consensus: Orchestrated an 8-agent pipeline that ingests signals, calculates risk, queries Gemini, validates policy, and prepares an executable intervention in under 1.5 seconds.
  • Zero-Hallucination Safety Gate: Built a bulletproof deterministic policy verification system that enterprise security and compliance teams can trust.
  • End-to-End Live Google Calendar Execution: Successfully implemented real-time, bi-directional Google Calendar sync that dynamically converts meetings into focus blocks with immediate visual feedback.
  • Full Serverless Deployment on Google Cloud Run: Shipped a containerized, production-ready full-stack application backed by Google Cloud Firestore and Secret Manager.

What we learned

  • Agent Specialization Beats Monolithic Prompts: Breaking down complex operations into single-responsibility micro-agents drastically reduced token latency, improved reasoning precision, and made the entire system modular and testable.
  • Grounding Builds Human Trust: Engineering managers embrace AI automation when every recommendation is transparently grounded with verifiable PR links, exact timestamps, and measurable time savings.
  • The Power of Google Cloud Serverless: Combining Gemini 3.5 Flash with Cloud Run and Firestore allowed us to build an enterprise-grade agentic system with virtually zero DevOps maintenance.

What's next for FlowPilot: AI Agent for Engineering Team Productivity

  • Interactive Slack & Discord Bot: Direct interactive notifications allowing managers to approve calendar conversions with a single Slack command (/flowpilot approve).
  • Predictive Developer Burnout Alerts: Analyzing cross-sprint communication patterns and after-hours PR reviews to alert managers before critical team members burn out.
  • Autonomous Jira Subtask Decomposition: Automatically breaking down large, blocked pull requests into smaller, parallelizable review tickets to eliminate long review queues.

Built With

  • docker
  • fastapi
  • google-calendar-api
  • google-cloud-firestore
  • google-cloud-run
  • google-gemini-3.5-flash
  • google-secret-manager
  • javascript
  • python-3.11
  • react-18
  • tailwind
  • vite
Share this project:

Updates

Submission history