Araca Insights

An offline desktop app for semiconductor wafer-polishing (CMP) test data. It imports instrument files, plays back runs, compares them and exports reports.

Creator and product owner Feb 2026 - Present
PySide6 Dash scikit-learn FLAML LangChain Ollama

Araca is a semiconductor CMP (chemical-mechanical planarization) equipment and services company. CMP is the step that polishes wafers flat. Every run on a test stand leaves instrument files behind, and someone then has to read those files, compare the runs, plot the correlations and build the report. Araca estimates that work at 35-95 hours per project when done by hand.

The company needed one tool for that work. It had to import the instrument files, play back each run, compare and correlate runs, and export the reports to Excel and PowerPoint. So I built Araca Insights from scratch.

A desktop app that runs offline

I wrote Araca Insights as a desktop app in Python 3.11. Three decisions shaped it.

Dash and Plotly inside a PySide6 window

The charts and controls are a Dash and Plotly interface, embedded in a PySide6 desktop window, and the whole app runs offline. Customer data stays on site. I picked PySide6 for its LGPL license, which lets the app ship as closed-source commercial software.

WHY → works offline in the lab, and can be sold as closed-source software
Portable project folder

Each project's data lives in a portable project folder, so moving a project to another machine means moving that folder.

WHY → a project travels as a folder
A local model for the agent

I built the AI agent (LangChain, Ollama) on a model that runs on the customer's own machine. It answers plain-language questions about process data and draws charts.

WHY → customer data never leaves the site

What the app does

Analysis Views
Imports instrument files and plays back each run. Correlation views plot the Stribeck curve, COF vs PV, pad temperature vs PV, Arrhenius, Preston, WIWNU vs PV and power density. The Compare view reports runs as ratios and differences with 95% intervals.
Prediction
Process-prediction models trained with scikit-learn and FLAML. AutoML searches nine estimator types, with cross-validation. Where the model allows it, each prediction comes with an uncertainty estimate from tree-ensemble variance.
Report Export
Reports are generated automatically and exported to Excel and PowerPoint, using XlsxWriter and python-pptx.
AI Agent
An on-premises LLM agent (LangChain, Ollama), included in the Advanced edition. It answers plain-language questions about process data and draws charts. Customer data never leaves the site.
Live Acquisition
Reads live data from National Instruments DAQ hardware through nidaqmx. It reads the inputs only. It doesn't control the polisher.
Licensed Builds
Windows builds compiled with Nuitka and installed with Inno Setup. License files are Ed25519-signed and bound to one machine. Two editions, Basic and Advanced: Advanced adds run comparison, the AI agent and prediction.

What shipped

35-95h Hours of Manual Work It Replaces per Project (Araca's estimate)
9 Estimator Types Searched by AutoML
On-prem LLM Agent (data stays on site)
Licensed Shipped as a Commercial Product

Building Insights as an offline desktop app was the right call for labs. It doesn't need a connection, and customer data stays on site. The agent follows the same rule. It uses a local model served through Ollama, so a plain-language question about process data gets an answer and a chart, and nothing leaves the site.

Shipping it as a product was a separate job. It took compiled Windows builds, an installer, signed license files bound to one machine, and separate Basic and Advanced editions, and I led the negotiation of the customer software license agreement. Insights is now licensed commercial software. I own the product line: roadmap, architecture, releases, licensing and first-line support.