DATA ENGINE · EN / FR · LIVE

AURA

Universal Data Engine

Any file in. A branded, bilingual report out. Five minutes, not five hours.

Drop a CSV, XLSX, JSON or Parquet. AURA infers what each column actually is, cleans it in eight steps, builds an archetype-aware explore, answers grounded questions in EN or FR, and exports a branded PDF.

Drop a file in ↗View source
55/55
Tests green
4
File formats in
8
Cleaning steps
2
AI providers live
WHAT I BUILT

It reads the data. It doesn't guess.

🧬01
Semantic Column-Role Inference
AURA does not just read dtypes. It decides what each column is FOR — identifier, measure, dimension or temporal — and every downstream step reads that decision. Get this wrong and the rest of the pipeline is confidently wrong too.
🔁02
Four-Provider AI Cascade
Gemini → Groq live, with OpenAI and Claude pluggable behind the same interface and silent automatic fallback. A rate limit or an outage at one provider does not reach the user as an error; it reaches them as a slightly slower answer.
🔒03
Grounded Chat, Not Free Association
Pre-computed findings are injected into the prompt as AUTHORITATIVE. The model cites numbers from the actual dataframe instead of inventing plausible ones — the difference between a demo and a tool.
🌍04
Bilingual All the Way Down
EN/FR is not a UI string swap. The exported PDF follows the locale through its labels, its AI executive summary, its insights, and its number and date formatting.
THE PIPELINE

Ingest → Infer → Clean → Explore → Report.

Every stage reads the one before it. The inference is the load-bearing step.

01
Ingest
CSV, XLSX, JSON, Parquet — parsed into one typed frame.
02
Infer roles
Identifier, measure, dimension, temporal — per column.
03
Clean
Eight steps. Dedup skips identifier columns, so near-duplicates are actually caught.
04
Explore
Charts chosen from the archetype, not from the dtype.
05
Report
Playwright + Jinja2, server-side SVG charts, EN or FR.
COLUMN-ROLE INFERENCE

An ID is not a number.

Most tools see customer_id as an integer and cheerfully offer you its mean. AURA classifies each column by what it is for, and every later stage — dedup, charting, the AI summary — reads that classification.

✓Identifiers excluded from statistics and from dedup keys
✓Measures get distributions; dimensions get frequencies
✓Temporal columns unlock trend charts automatically
orders_2024.csv · 6 columns
order_id
High cardinality, no meaningful order
IDENTIFIER
order_date
Parses as a date, monotonic-ish
TEMPORAL
region
Low cardinality, repeats
DIMENSION
revenue
Continuous, aggregates meaningfully
MEASURE
units
Discrete but additive
MEASURE
customer_id
Numeric — and averaging it is nonsense
IDENTIFIER
TECH STACK

55 tests green. Both languages covered.

FastAPINext.jsTypeScriptPandasGeminiGroqOpenAIClaude APIPlaywrightJinja2SSE streamingPytest + 55-check E2E
UNIVERSAL DATA ENGINE

Bring a messy file.
Leave with a report.

Upload any CSV, XLSX, JSON or Parquet and watch the roles get inferred, the cleaning run, and the branded PDF come out the other side — in English or French.

Open AURA ↗Source on GitHub