Inspiration
I grew up watching small business owners in Nigeria manage everything by memory and instinct. No tools, no analyst, no way to know when something was quietly going wrong until it was too late.
Most data analysts in Nigeria only work with large companies — the kind that can afford retainers and enterprise BI tools. The suya spot owner running three branches across Lagos, the boutique owner managing two locations in Abuja — they are left out entirely. They have no visibility into their own numbers. But they all have WhatsApp open all day.
That gap is what DataWatch is built to close.
What it does
DataWatch is an autonomous business intelligence agent for small business owners. It connects to a business's Google Sheets sales data, learns from 90 days of historical patterns across three branches — Ikeja, Surulere, and Lekki — and uses Gemini 3.5 Flash to reason about what is genuinely unusual.
When a severe anomaly is detected, the business owner receives an instant WhatsApp alert with the branch, the hour, how far below normal it is, and a likely cause. Every Monday at 8am, a full weekly summary is sent automatically. No dashboard to log into. No analyst to hire. Just WhatsApp.
How we built it
The agent is built on the Google Cloud stack. Google ADK powers the autonomous agent loop with four tools: analyze_sales, generate_chart, generate_report, and remember. Gemini 3.5 Flash provides contextual anomaly reasoning — it understands peak hours, day-of-week patterns, and owner-defined exceptions stored in memory. Google Sheets API serves as the lightweight data layer. Cloud Run hosts the deployment endpoint and Compute Engine runs the agent runtime and cron jobs.
The agent runs on two autonomous schedules — an hourly severe anomaly check and a Monday 8am weekly report — with no human intervention required after setup.
Challenges we ran into
Model availability was the first wall. Several Gemini model versions had been deprecated by the time I tried to use them and had to be identified dynamically.
Google Cloud authentication took significant time. Service account key creation was blocked by an org policy that had to be manually overridden.
Prompt engineering was harder than expected. Getting Gemini to reason contextually — rather than just flagging numbers below a threshold — required careful construction of the business profile prompt and multiple iterations.
Scope discipline was the hardest challenge. Knowing what to cut and what to keep, with a hard deadline, is a skill in itself.
Accomplishments that we're proud of
The moment Gemini flagged a ₦300 sales figure at Ikeja during peak lunch hour — not because of a hardcoded rule, but because it understood that 12pm on a Sunday should generate over ₦60,000 — and described it as a likely POS failure or power outage. That is genuine contextual reasoning, not a script.
Building a complete autonomous agent — from data ingestion to WhatsApp delivery — as a solo project in two weeks, while learning Google ADK from scratch.
What we learned
Building DataWatch taught me that the most powerful thing an AI agent can do is not answer questions — it is take action without being asked. I learned the real difference between a script and an agent. I also learned that prompt engineering is not about clever tricks — it is about giving the model the right context to reason well.
What's next for DataWatch
- Connecting to a Nigerian public holidays API so the agent automatically accounts for expected low-sales days
- Multi-branch expansion beyond three locations
- Support for other messaging platforms — Telegram and SMS for owners without WhatsApp
- A proper owner onboarding flow so any business can connect their own Google Sheet and go live in minutes
Log in or sign up for Devpost to join the conversation.