$ cat case-studies/lumiq.md · 03 / 04
LUMIQ
$ what it had to do
Thejob.
Most people with a spreadsheet can’t write SQL, and an AI chat on its own will happily make numbers up. LUMIQ does the maths in code, right in the browser, and only lets Llama 3.3 70B (on Groq) write the sentences around figures that are already checked. Rows never leave your machine.
$ trace --request
How itworks.
- 01Uploada CSV, read in the browser
- 02Aska question in plain English
- 03Computethe numbers, in code, checked
- 04Explainthe answer in words, by the model
- Natural-language filtering
- Crystal Ball forecasting with R² confidence bands
- Anomaly detection by Z-score
- Pearson correlation heatmap
- Decision Brief generator
- Scenario Forge
- Data DNA profiler
$ git log --grep=decision
The callsthat mattered.
- decision.1Numbers from code, words from AI
The model never does arithmetic. It is handed figures that code has computed and checked, and writes the explanation around them.
- decision.2No backend at all
The data is read and analysed in the browser, so rows never leave the visitor’s machine. Only the question and the computed figures go to the model.
- decision.3Every chart drawn by hand
All the charts are pure SVG, written for LUMIQ, with no chart or UI library in between.
built with
- React 18
- Vite
- Groq API
- Llama 3.3 70B
- Pure SVG