Skip to content

yajatns/Feb27Hackathon

Repository files navigation

backoffice.ai

One AI that runs your entire back office. You talk. It handles the tools.

🏆 Built at the Render x Friends Hackathon — Feb 27, 2026

🌐 Live Demo · API Docs


What is backoffice.ai?

Founders spend 40% of their time on operational busywork — hiring, compliance, payroll, IT provisioning. All manual. All fragmented across dozens of tools.

backoffice.ai replaces that with a team of autonomous AI agents. You tell it to hire someone. It handles everything — salary benchmarking, policy compliance, benefits enrollment, IT provisioning, system integrations — and explains every decision it makes.

How It Works

CEO: "Hire Sarah Chen as Senior Engineer, $180K, San Francisco"

🤖 Orchestrator → Delegates to 5 specialist agents:

👩‍💼 Maya (HR)       → Queries Senso for salary bands & onboarding policy
📊 Sam (Finance)    → Researches market rates via Tavily ($131K-$204K range)
⚖️ Compliance       → Checks labor regulations & internal policies
💻 Alex (IT)        → Provisions accounts via Yutori portal automation
🔗 Aria (Integrations) → Syncs to Notion/Salesforce/Stripe via Airbyte

✅ Final Decision: "APPROVE — salary within band, all checks passed"

Every agent is a real LLM with tools (not scripted functions). They reason independently, call real APIs, and explain their logic.


Key Features

🤖 Autonomous Multi-Agent Pipeline

  • Orchestrator decides which specialists to invoke using OpenRouter function calling
  • Each specialist has its own system prompt, tools, and multi-turn reasoning loop
  • Agents run in parallel where possible, sequentially when there are dependencies

🧠 Self-Improvement Loop

When a human overrides an agent's decision:

  1. Override recorded as LEARNED edge in Neo4j
  2. Local policy store updated immediately
  3. Cron detects patterns (3+ overrides same direction)
  4. Generates updated policy → next hire gets smarter

🕸️ System of Reasoning (Neo4j)

Every delegation, tool call, completion, and override is traced as a knowledge graph. Full auditability — when the CEO asks "why did you offer $195K?", we show the exact reasoning chain.

🚨 Red Flag Detection

Sam (Finance) cross-references salary offers against real market data. Lowball offers (>15% below market) are flagged as CRITICAL and blocked from proceeding.

🔗 600+ System Integrations

Aria (Airbyte agent) discovers and connects to any system — Notion, Salesforce, Stripe, Jira, GitHub, and 600+ more via PyAirbyte.


Sponsor Integrations

Sponsor How We Use It Agent
OpenRouter Powers ALL agent LLM calls (Claude 3.5 Sonnet with function calling) All
Senso Policy knowledge base — salary bands, compliance, benefits. Self-improvement target Maya, Compliance
Neo4j Aura System of Reasoning — traces every delegation, tool call, override as a graph All
Tavily Real-time salary benchmarking from salary.com, ZipRecruiter, levels.fyi Sam
Yutori Portal automation for benefits enrollment & account provisioning Alex
Reka Vision API for document analysis and video compliance auditing Query endpoint
Airbyte Universal connector — 600+ systems, connector discovery & data sync Aria
Render Infrastructure — API (FastAPI), Dashboard (Next.js), PostgreSQL

Architecture

┌─────────────────────────────────────────────────────┐
│                    CEO / User                        │
└──────────────────────┬──────────────────────────────┘
                       │
              ┌────────▼────────┐
              │   Orchestrator  │ ← OpenRouter (Claude 3.5 Sonnet)
              │   (LLM Agent)   │
              └───┬──┬──┬──┬──┬┘
                  │  │  │  │  │
         ┌────────┘  │  │  │  └────────┐
         ▼           ▼  ▼  ▼           ▼
    ┌─────────┐ ┌────┐ ┌──┐ ┌────┐ ┌──────┐
    │  Maya   │ │Sam │ │⚖️│ │Alex│ │ Aria │
    │  (HR)   │ │(Fin)│ │  │ │(IT)│ │(Int) │
    └────┬────┘ └──┬─┘ └┬─┘ └──┬─┘ └───┬──┘
         │         │    │      │        │
    ┌────▼────┐ ┌──▼──┐ │  ┌───▼──┐ ┌───▼────┐
    │  Senso  │ │Tavily│ │  │Yutori│ │Airbyte │
    │(Policy) │ │(Mkt) │ │  │(Auto)│ │(600+)  │
    └─────────┘ └─────┘ │  └──────┘ └────────┘
                    ┌────▼────┐
                    │  Senso  │
                    │+ Tavily │
                    └─────────┘
                         │
              ┌──────────▼──────────┐
              │      Neo4j Aura     │
              │  (Reasoning Graph)  │
              └──────────┬──────────┘
                         │
              ┌──────────▼──────────┐
              │   Self-Improvement  │
              │  Cron → Senso Upload│
              └─────────────────────┘

Demo

AlexSaaS is our demo customer — a 50-person SaaS startup using backoffice.ai.

Try It Live

  1. Dashboardhttps://backoffice-dashboard-kqya.onrender.com
  2. New Hire → Submit an employee → Watch 5 agents reason autonomously
  3. Pipeline View → See every agent's tools, reasoning, and decisions
  4. Neo4j Graph → Visual trace of the entire reasoning chain
  5. Query → Ask anything: "What's our salary band for engineers in SF?"
  6. API Docshttps://backoffice-api-ep7k.onrender.com/docs

Tech Stack

Backend: Python, FastAPI, SQLAlchemy, asyncpg, PostgreSQL, httpx, Pydantic Frontend: Next.js 16, TypeScript, Tailwind CSS, vis-network AI: OpenRouter (Claude 3.5 Sonnet), function calling, multi-turn agent loops Data: Neo4j Aura (graph), PostgreSQL on Render (persistence) APIs: Senso, Tavily, Yutori, Reka, Airbyte (PyAirbyte) Infra: Render (3 services — Web API, Static Site, PostgreSQL)


Project Structure

├── backend/
│   ├── agents/           # AI agents (orchestrator, HR, finance, compliance, IT, airbyte)
│   │   ├── base.py       # BaseAgent — LLM reasoning loop with tools
│   │   ├── orchestrator.py # Orchestrator — delegates via function calling
│   │   ├── hr_agent.py   # Maya — Senso policy search
│   │   ├── finance_agent.py # Sam — Tavily salary benchmarking
│   │   ├── compliance_agent.py # Compliance — regulations + internal policy
│   │   ├── it_agent.py   # Alex — Yutori portal automation
│   │   └── airbyte_agent.py # Aria — 600+ connector discovery
│   ├── integrations/     # API clients (OpenRouter, Senso, Tavily, Neo4j, Yutori, Reka, Airbyte)
│   ├── models/           # SQLAlchemy models + Pydantic schemas
│   ├── routes/           # FastAPI routes (hire, query, graph, override, chat, crons, airbyte)
│   └── main.py           # App entry point
├── frontend/
│   ├── app/              # Next.js app router (dashboard, hire, graph, query)
│   ├── components/       # React components (HireForm, PipelineView, GraphViewer, etc.)
│   └── lib/              # API client + WebSocket
└── render.yaml           # Render Blueprint (Infrastructure as Code)

Setup

# Clone
git clone https://github.com/yajatns/Feb27Hackathon.git
cd Feb27Hackathon

# Backend
cp .env.example .env  # Fill in API keys
cd backend
pip install -r requirements.txt
uvicorn main:app --reload

# Frontend
cd frontend
npm install
npm run dev

Environment Variables

OPENROUTER_API_KEY=     # LLM calls (all agents)
SENSO_API_KEY=          # Policy knowledge base
NEO4J_URI=              # Graph database
NEO4J_USER=             # Graph auth
NEO4J_PASSWORD=         # Graph auth
TAVILY_API_KEY=         # Market research
YUTORI_API_KEY=         # Portal automation
REKA_API_KEY=           # Vision API
DATABASE_URL=           # PostgreSQL connection string

Team

Electrons in a Box 🔌

Member Role
Nag (@nagaconda) Product & Strategy
Yajat (@yajatns) Engineering Lead
Chhotu 🤖 Frontend & Demo
Cheenu 🐿️ Backend & API

Yes, half our team is AI agents. That's the point.


License

MIT

About

autonomous agents hackathon repo

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors