From market research to a live pilot in four days
This is the first post of a build-in-public journal. The deal: we show what we ship, what breaks, and what it costs — with real numbers where they're ours to share.
The itch
Three different businesses asked us the same question in one month, each in their own words: "I have a CRM, a task tracker, ad campaigns, spreadsheets — why can I still not see what's going on?"
A restaurant chain with a webshop, bots and custom tooling. A trading company running on ClickUp, Zoho and Excel. A team that literally asked for "a management dashboard on top of the systems we already have."
When the market repeats itself three times, you stop consulting and start building.
Day 1: research before code
Before writing a line, we ran a deep market sweep: four camps of competitors (all-in-one suites, vertical POS platforms, AI copilots, BI aggregators), pricing, and the graveyard — why all-in-one products die.
Three findings shaped everything:
- Nobody combines one data core + role-based dashboards + a proactive AI layer. Suites have modules, BI tools have charts, copilots have chat. The intersection is empty.
- An AI layer built on someone else's API is a rented foundation. Platforms cut off integrations exactly when you become valuable. So connectors are an on-ramp, not the foundation — the system of record must be ours.
- SMBs buy consolidation when value shows up in the first week, setup doesn't require "hiring a guy," and the data can leave anytime.
Days 2–3: the boring architecture that works
One TypeScript monorepo. One Postgres — with row-level security doing multi-tenant isolation at the database, not in app code. Connectors pull raw records append-only; normalizers build a canonical model (orders, payments, invoices, parties — with lineage on every number). A metrics registry defined in code materializes daily values; statistical detectors (z-score, week-over-week, thresholds) find anomalies before any AI is involved.
The AI layer never does arithmetic. It reads the metrics, the anomalies and the business profile, and writes insights with evidence links — every claim traceable to a number. No key configured? The system degrades honestly to detector-driven insights, clearly labeled.
Day 4: real data or it didn't happen
We pointed the pipeline at a real store: 1,623 live orders flowed through, zero normalization errors. The first nightly AI run produced six insights — and our favorite was the AI complaining about us: "ad spend, margins and delivery-time data are not connected yet — half of the pilot's goals aren't measurable." Correct. That honesty is the product.
It also caught a real data-quality gap we didn't know about: most orders carried no city, because the checkout kept the city in a cookie instead of the order. One future one-line fix — and location analytics lights up. A dashboard shows you numbers; a system tells you why you can't trust them yet.
What it costs
- Server: €5.49/month (2 vCPU, EU). Postgres, web, worker, TLS, daily backups.
- AI: roughly $15–30/month per active tenant — nightly batch analysis plus Q&A, with prompt caching doing the heavy lifting.
- Team: one engineer, pair-building with AI. The research, the code, the deploy, this post.
What's next
Connect the money side (bank statements) and advertising, then the morning Telegram digest. And the first paying pilots — that's the part you can't automate.
If your business runs on five tools and zero visibility — that's exactly the itch we're scratching. The audit is free: t.me/agtsurkan.