Skip to main content
biznesam.ai
· 4 min read AIAnalyticsAutomationDashboard

A Business Analytics Tool That Doesn't Just Show Data — It Thinks and Recommends

6 data sources on one screen, built-in AI analyst, and a Karpathy autoresearch cycle that automatically optimizes business metrics. How we built an analytics system that finds opportunities on its own.

FastAPIPythonClaude APIMeta APIGoogle Ads APIPlausible
A Business Analytics Tool That Doesn't Just Show Data — It Thinks and Recommends

Why Yet Another Dashboard?

The world doesn’t lack dashboards. Web analytics, ad platforms, email tools, CRM reports — every service offers its own. The problem isn’t lack of data. The problem is data fragmentation.

A service company’s executive opened 6 tabs every morning, looked at 6 different charts, and tried to piece together the full picture in their head. “Why did conversion drop this month?” they asked in a meeting. No one could answer — because the answer was scattered across 3 different systems.

All Data on One Screen

First step — combine 6 data sources into one dashboard:

  • Sales data — deal stages, stage-by-stage conversion, average deal value, stale deal detection
  • Paid advertising channels — spend, return, conversions by campaign
  • Social media channels — reach, engagement, follower dynamics
  • Website analytics — visits, bounce rate, session duration, traffic sources
  • Email marketing — opens, clicks, unsubscribes, engagement trends
  • Internal company data — transaction volume, seasonality, forecasts

Traditionally, each API connection takes 3–5 business days. Six connections — 4–6 weeks just for integrations. We did it in a day.

AI Analyst: Not Charts, But Answers

The dashboard includes a built-in AI analyst — not a chatbot, but a full analytics partner with access to all data.

What it does:

Ask a question in natural language — “How did we do this week?” — and receive a data-driven answer. Not “sales were €X” but “sales grew 12% compared to last week, mainly due to campaign Y, but email conversion dropped — recommendation: review subject lines.”

Automatic anomaly detection. If costs spike in a channel or conversion drops — the system alerts proactively, without waiting for someone to notice.

Smart questions. AI doesn’t just answer — it suggests questions worth asking. “Did you notice that Thursday emails show 40% higher open rates than Monday ones?”

Karpathy Autoresearch: AI That Optimizes Itself

The second part of the system is inspired by Andrej Karpathy’s autoresearch concept — automatic iterative optimization.

Karpathy used it for automating machine learning experiments. We brought it to business:

The cycle:

  1. Modify — AI generates a new variation (email subject, meta description, follow-up message, ad copy)
  2. Measure — the system measures results via integrated data sources (open rate, bounce rate, conversion, ROAS)
  3. Keep/Revert — if the improvement is statistically significant — keep it. If not — revert
  4. Repeat — the cycle continues automatically

The key difference from ML experiments: measurement period. GPUs give results in 5 minutes. Real business metrics take 7–14 days. But the principle is identical — and it works.

The Insight That Changed Budget Decisions

Analyzing historical sales data, the AI analyst identified something no one had noticed:

One paid channel showed a 79% non-response rate — compared to 50% average from other channels.

That means out of every 10 clients from that channel, only 2 ever responded. The other 8 — budget wasted.

This insight significantly impacted budget allocation and efficiency.

Architecture

The dashboard runs on a FastAPI + pandas backend:

  • Pandas pipeline for data cleaning, filtering, and aggregation
  • Department filters (sales, marketing, management)
  • Date range selection (week, month, quarter, year)
  • Real-time data refresh

The AI chatbot uses Claude API with full data context — meaning every question is answered with all available information, not just one data source.

What We Learned

Pandas is underrated for business analytics. Data cleaning, filtering, aggregation — all in one pipeline. No separate data warehouse needed to start getting value.

AI + data > AI + text. AI chatbots answering from “knowledge” are mediocre. AI chatbots with access to real data are transformative. The difference is enormous.

Management wants answers, not charts. Nobody looks at charts for fun. Management wants to know: “are we doing well?” and “what should we change?” The AI analyst answers exactly these questions.

Are your business data scattered across different systems? Let’s talk — free consultation about your opportunities.