FlowForge

The Inspiration Every year, organizations collectively leak billions of dollars due to broken, delayed, or fragmented employee onboarding processes. While researching this problem, we discovered several devastating statistics from the Society for Human Resource Management (SHRM) and Gallup:

$11,000+ is the average organizational cost of a single bad onboarding experience (due to lost productivity and early attrition). 45% of new hires quit within their first year when subjected to disorganized, check-list-driven onboarding. 8 to 10 Hours of manual labor is drained from HR managers per new hire just to coordinate tasks. We realized that traditional HRIS platforms (like Workday or BambooHR) are passive databases—they don't actually do the work. HR personnel are stuck manually copying email addresses, provisioning Slack channels, requesting GitHub organization access, scheduling manager 1:1s, and chasing down signatures on PDF forms.

We asked ourselves: What if we built an autonomous, self-healing team of AI agents that could coordinate these tasks, reason about optimal mentor pairings, handle transient network failures, and complete the entire process in under 20 minutes?

This question inspired FlowForge Onboard.

🏗️ How We Built It We designed FlowForge Onboard as a hybrid architecture built for reliability, real-time feedback, and high efficiency:

1. The Multi-Agent Orchestration Engine (Backend) The backend is built in Python 3.11 using FastAPI and native asynchronous execution via asyncio. The architecture implements a directed acyclic graph (DAG) model where 6 specialized agents communicate asynchronously over a typed message bus:

Orchestrator Agent: Coordinates task dispatch, state machines, and dynamically handles failures. Account Provisioner 🔐: Simulates account creations in Slack, GitHub, Jira, and Google Workspace. Document Validator 📋: Chases down paperwork requirements (Offer letters, I-9, Direct Deposit info). Buddy Matcher 🤝: The core reasoning agent that pairs new hires with mentors. Schedule Coordinator 📅: Auto-books calendar events based on manager/new-hire availability. Anomaly Detector 🔍: Monitors processing times, tracks repeated failures, and sends Slack escalations.

1. The Buddy Matching Algorithm (The Math) Email me for the math
2. The Visual Dashboard (Frontend) The frontend dashboard is built using Next.js 15, compiled to a static Single Page Application (SPA), and deployed to Google Cloud Run. It connects to the backend microservice over secure WebSockets (wss://) to stream live agent updates. We designed the UI using a premium Kyoto Cyber-Minimalist aesthetic, utilizing deep Sumi-ink matte black backgrounds (#080809), hairline borders, gold leaf/bronze accent tokens (#c5a880), and Hanko vermilion status colors (#ff5e54).

🚧 Challenges We Faced Self-Correction & Resilience: Building an agent system for the "happy path" is easy. Designing for real-world enterprise infrastructure—where APIs rate-limit, databases lock, and networks timeout—is difficult. We had to build a complex state-tracking engine in the Orchestrator to distinguish between transient errors (which it automatically retries using exponential backoff: T retry ​ =2 attempt seconds) and fatal errors (which trigger a Slack escalation message containing interactive blocks). CORS and Static Hosting Mismatches: During deployment, we had to work around permission issues in Google Cloud/Firebase APIs. We successfully solved this by hosting both the frontend static files (using npx serve -s inside a lightweight Node/Alpine container) and the backend FastAPI service directly on Google Cloud Run, configuring the backend with a wildcard CORS regex https?://.* to dynamically allow cross-origin WebSocket and fetch operations. Hydration Sorting Errors: Next.js pre-rendered locale-specific timestamps on the server side that mismatched the client browser's timezone, causing React rendering failures. We implemented a timezone-defensive sorting algorithm and a client-side mounting guard to resolve this. 🎓 What We Learned AI Orchestration is a Distributed Systems Problem: AI agent swarms behave exactly like microservices. To keep them reliable, we had to rely on classic systems engineering patterns: typed message envelopes, state machines, append-only SQLite logs configured in Write-Ahead Logging (WAL) mode, and health check routes. Autonomy Requires Observability: HR teams will never trust AI if it operates as a black box. The most important part of FlowForge isn't that it pairing employees autonomously, but that it writes its entire decision-making process into a printable audit trail and updates the dashboard in real-time. Specialization makes each node testable and easily optimized. 🔮 What's Next for FlowForge Onboard OAuth Integration: Swap our simulated provisioning wrappers for production-grade OAuth pipelines into GitHub, Google, Slack, and Jira APIs. LLM Document Intelligence: Integrate Vision Language Models (VLMs) to read and audit photo uploads of onboarding documents (verifying signatures and dates). SaaS Multi-Tenancy: Enable zero-config deployment so organizations can connect their Workday database to FlowForge with one click.

Built With

• all-managed-securely-via-environment-variable-injection.-version-control-is-maintained-using-git
• alongside-pydantic-v2-for-strict-input-data-validation-and-model-serialization.-real-time-updates-are-streamed-directly-to-the-user's-browser-using-native-websockets
• and
• and-highly-performant-stack.-the-core-codebase-is-written-in-python-3.11-for-the-backend-orchestration-layer-and-typescript-for-the-frontend-application
• and-jira-api
• and-relies-heavily-on-python?s-native-asyncio-library-to-handle-concurrent-multi-agent-executions
• compiled-to-a-static-single-page-application-(spa).-both-the-frontend-and-backend-are-containerized-using-docker-and-deployed-serverlessly-to-google-cloud-platform-(gcp)-via-google-cloud-run
• container
• custom-vanilla-css-to-achieve-a-kyoto-cyber-minimalist-design-without-relying-on-bloated-external-utility-libraries.-the-backend-microservice-is-built-using-fastapi-as-the-api-framework
• file
• flowforge-onboard-was-developed-using-a-modern
• frontend
• google-workspace-api
• hosted-on-github
• in
• is
• local
• managed
• npx
• powered-by-the-uvicorn-asgi-server
• serving
• slack-webhooks
• static
• styled-using-clean
• the
• type-safe
• via
• we-utilized-sqlite-configured-in-write-ahead-logging-(wal)-mode-with-a-customized-busy-timeout-via-aiosqlite-to-guarantee-high-concurrency-database-safety.-the-frontend-is-built-on-next.js-15-and-react-19
• while-external-api-integrations-are-managed-asynchronously-using-the-httpx-client-library.-for-data-persistence
• with-container-compilation-automated-remotely-through-google-cloud-build-and-stored-in-google-artifact-registry.-the-system-integrates-simulated-and-mocked-wrappers-for-the-github-api