backoffice.ai

What Inspired This?

We spoke to a couple of founder friends and kept hearing the same thing: "I lose entire days to operational busywork." One friend said hiring a single engineer took him 15 hours — employee record, payroll setup, IT provisioning, updating systems. All manual. All fragmented.

We did more research and realized: this is universal. Founders spend 40% of their time on back-office work instead of building their product. We thought: What if AI agents could just... handle this?

How We Built It

We created a team of autonomous AI agents — each a specialist in one domain of back-office operations. The CEO talks to one orchestrator, and it delegates to the right specialist. Every agent reasons independently using OpenRouter (Claude 3.5 Sonnet with function calling), deciding which tools to invoke and why.

The magic is in how we connected the sponsors to solve real problems:

Senso became our policy brain — we loaded AlexSaaS's salary bands, compliance checklists, and benefits policies into Senso's knowledge base. Before any agent makes a decision, it queries Senso to ground its reasoning in company truth. More importantly, when a human overrides a decision, that correction feeds back into Senso — the agents literally learn from every correction and get smarter over time. That's our self-improvement loop.

Tavily gives our Finance agent (Sam) real-time market intelligence. When Sam benchmarks a salary offer, he pulls live data from salary.com, ZipRecruiter, and levels.fyi — not cached data, not estimates. Real numbers. This is how we caught that a $140K Senior Engineer offer in SF was 18% below market.

Reka provides vision capabilities for document analysis — parsing resumes, verifying IDs, and auditing training videos for compliance. It sees what text-only models can't. [ In Progress ]

Airbyte is our universal connector layer. With 600+ connectors, our integration agent (Aria) can sync employee data to any system the customer already uses — Notion, Salesforce, Stripe, Jira. No new tools to learn. [ In Progress ]

Neo4j Aura is our System of Reasoning — every delegation, every tool call, every human override is traced as a graph. When the CEO asks "why did you offer $195K?", we can show the exact chain: Senso policy → Tavily benchmark → human override → learned policy update. Full auditability.

Render hosts everything — our FastAPI backend, Next.js dashboard, and PostgreSQL database. Three services, zero DevOps.

The Challenges

Cross-system integration — Different APIs, different auth, different data models. Had to build abstraction layers so agents think in business tasks, not API calls.

Making agents truly autonomous — We built every agent into an LLM with its own tools and reasoning loop. The orchestrator doesn't follow a script — it decides which specialists to call based on the request.

The self-improvement loop —: hire → human corrects → policy updates → next hire is smarter. That closed loop is what makes this more than a chatbot.

Trust & safety — The Neo4j graph means every decision is auditable.

Built With

• airbyte-(pyairbyte-+-agent-connectors)
• fastapi
• httpx
• neo4j-aura
• next.js
• notion-api
• openrouter-(claude-3.5-sonnet)
• postgresql
• pydantic
• python
• reka-vision-api
• render-(web-services-+-postgresql)
• senso-api
• sqlalchemy
• tailwind-css
• tavily-api
• typescript
• yutori-api