Case study · personal project · 2026

AI Business Operations Hub

A working simulation of a company run by one human and four AI agents: lead qualification, customer service with citations, invoice follow-up, and task assignment — all behind one approval queue and one command center.

Role design + build Stack FastAPI · PostgreSQL · pgvector · n8n Model DeepSeek via gateway Status running, private demo

Agents

n8n workflows

Unit tests

Run errors / 15

The problem

A 20-person digital agency loses hours to work that repeats: reading inbound leads, writing the same follow-up, chasing invoices, turning meeting notes into tasks. The information exists, but nobody has one place to see it — so the founder asks the team instead of reading a dashboard.

The goal was not another chatbot. It was a small operating system where agents do the repetitive part, a human keeps every decision that carries money or reputation, and every action is logged well enough to audit.

What runs

Sales agentScores a lead 0–100, picks a band, drafts the follow-up. Outbound message always waits for approval.

Customer service agentAnswers only from retrieved documents, cites sources, escalates when confidence is low or the message smells like a complaint.

Finance agentFinds overdue invoices, applies the payment policy, drafts the reminder. Sending it needs a human.

Operations agentTurns a meeting note into a task with owner and deadline; rejects deadlines in the past.

ChannelForm / webhook / email

Orchestrationn8n triggers the API

CoreFastAPI + agent + policy

ControlApproval queue → dashboard

Command center

Human-in-the-loop, by rule

Autonomous

Scoring, classification, CRM updates, drafting, internal alerts, routine task creation.

Needs approval

Any outbound message, discounts, invoice reminders, payments, strategic tasks, complaints — plus anything with confidence below 0.70.

Numbers that exist

Metric	Value	Source
Agent runs	15	agent_runs
Errors	0	agent_runs.status
Average latency	5.9 s	avg(latency_ms)
Approvals created	11	approval_queue
Retrieval hit@4	1.00	10 queries
Retrieval MRR	0.95	10 queries
Fallback band accuracy	6/6	scripts/eval_agents.py

Business impact (≈82% less drafting time per 100 leads) is an estimate built on an assumed 15-minute manual baseline. It is labelled as an estimate in the repository, not presented as a measured result.

30 · 60 · 90

30 days

Replace demo CRM with the real one
WhatsApp Business API for real intake
30-case golden set for agent accuracy

60 days

Approve from WhatsApp/Slack
Retry and circuit breaker at the gateway
Rate limiting and access audit

90 days

Reranker and hybrid search
Marketing and content agents
Weekly founder report

Honest limits

WhatsApp and email are simulated, not connected to production APIs. There is no demo video yet. Token cost per run is not recorded; the gateway does not return usage to the app. And the business-impact number is an estimate, not a measured deployment.

Open the private demo → Back to portfolio

The live demo is protected by HTTP basic auth; credentials are shared during the interview process.