AI analysis you can actually trust
- Language models are good at explaining data but can invent numbers.
- Messy real-world files: mixed types, inconsistent categories, missing values.
- Results had to be clear to people who don’t read raw statistics.
← All projects
Data · Full-stack
Autonomous Data Analyst
An AI-assisted analytics workspace: upload a dataset and get a data-quality report, findings and charts — backed by a deterministic profiling engine.
The first hour with a new dataset is always the same. This workspace does it for you.
Before any real analysis, every dataset needs the same groundwork: checking types, missing values, duplicates, outliers and distributions. Autonomous Data Analyst automates that first pass and presents it as a readable workspace — overview, data quality, insights and charts.
What made it hard, and how I approached it.
Deterministic numbers, readable findings
- A FastAPI profiling engine computes every metric deterministically — the same file always gives the same report.
- Issues are grouped by severity with the affected columns, so it is obvious what to fix first.
- Findings and charts are generated from the computed profile, keeping explanations grounded in real values.
Built around what people actually need.
Data quality
A canonical quality score with detected issues — missing values, inconsistencies, outliers.
Insights
Evidence-backed findings, each linked to the columns and values behind it.
Charts
Distributions and category breakdowns chosen automatically from the dataset’s structure.
Responsive
Responsive workspace
The workspace adapts to smaller screens with a collapsible sidebar, so reports can be checked on the go.
- Collapsible sidebar frees up the screen
- Quality score and findings stay above the fold
- Demo dataset works without uploading
What the profiling engine actually finds.
Both charts use the real output of the app’s FastAPI profiler, run on the deliberately messy demo dataset that ships with it.
Data quality
A score of 65, and the breakdown that explains it.
No missing values and no duplicate rows, but 14 issues: invalid quantities and percentages, an email column flagged as PII, and labels that mean the same thing written differently.
Source: profile_frame() in backend/main.py, run on the /api/demo dataset.
Consistency
Why consistency scores 55: one segment, two spellings.
“Enterprise” and “enterprise” count as different groups until the labels are standardised. Toggle to see the breakdown after cleaning.
Source: category_breakdown from the profiler output; the cleaned view merges case variants.
A closer look at the product.
Under the hood.
- FrontendVite + React + TypeScript workspace, deployed on Cloudflare Pages.
- Profiling enginePython / FastAPI service computing deterministic dataset profiles.
- Production APIDeployed as a Cloudflare Python Worker; local development uses FastAPI behind a Vite proxy.
- Demo modeA built-in demo dataset so visitors can explore without uploading anything.
ReactTypeScriptVitePythonFastAPICloudflare PagesCloudflare Workers
What I learned.
Separating “compute” from “explain” was the key decision. Keeping every number deterministic and only using AI-style summaries on top of it made the tool far more trustworthy.
Next project
EDEN Smart Terrarium
A smart dual-plant terrarium and its live dashboard — sensors, pumps, a fan and a grow light, built as a group project for the Interactive Systems course.
