SlothDesk：把培訓評估結果和參加者意見整理成可編輯報告

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady

What I Built

SlothDesk is a local AI assistant that turns training assessment results and participant feedback into an editable evaluation report.

The inspiration came from training coordination work. Finishing a training session does not finish the paperwork: someone still needs to compare assessment scores, read feedback, identify suggestions, and write the report.

SlothDesk makes that process easier:

• Upload pre-training and post-training results as CSV files.
• Match participants by ID and calculate score changes.
• Identify unmatched records.
• Generate a feedback summary using a locally running open-weight model.
• Review the evidence, edit the narrative, and download the report.

The person who gets to be a little lazier? The training coordinator who would otherwise assemble everything manually.

Demo

https://youtu.be/hllnCGUArz4

I tested the application on my Windows computer and successfully generated a summary using the local model.

The included synthetic dataset demonstrates:

• 4 matched participants
• 57.5% average pre-training score
• 80% average post-training score
• 22.5 percentage-point improvement
• Two unmatched IDs, listed separately and excluded from the comparison

The workflow is: sample data → calculated results → AI feedback summary → editable report → download.

Code

View SlothDesk on GitHub

The repository includes the application source, setup instructions, MIT licence, automated tests, and synthetic sample data.

How I Built It

SlothDesk uses Python’s standard library for its local server and CSV processing, with an interface built in HTML, CSS, and JavaScript.

At its AI core is Qwen2.5 1.5B, running locally through Ollama. The model reads feedback and drafts three sections:

• Strengths
• Improvements
• Follow-up actions

The prompt asks the model to reference feedback records using identifiers such as [F2]. Those source records are visible beside the workflow so the user can verify the draft.

Numerical analysis is handled directly by Python. Scores are normalised to percentages, and averages use only participants present in both assessment files. Duplicate IDs and invalid scores produce errors rather than guessed results.

The report downloads as standalone HTML, which can be opened in a browser and printed to PDF.

I built the project with assistance from ChatGPT/Codex. Testing on Windows exposed a text-encoding bug, which we fixed by explicitly reading the application files as UTF-8. Seven automated tests passed, and I confirmed that local AI summary generation worked on my computer.

Why Does Open Innovation Matter?

Training records and feedback can contain information that organisations prefer to keep on their own computers.

SlothDesk sends feedback to Ollama on the same machine instead of requiring a hosted AI API. Its interface also uses local assets, without external fonts or scripts.

Using an open-weight model gives me control over the workflow. I can inspect the prompt, change how summaries are written, and replace the model with another suitable local model.

There are no per-request model API charges. Running it still requires computing resources and electricity, and initial setup requires downloading the software and model.

For this project, the advantage over a hosted-only approach is practical: I can generate feedback summaries without making an external inference service part of the reporting process.

The open model performs a central task—turning free-text comments into a readable draft—while the user keeps responsibility for checking and approving the report.

The Lazy Payoff

I have tested the complete workflow with the supplied sample data, including local AI generation. I have not yet measured time savings on a real training programme.

The immediate benefit is having calculations, feedback evidence, drafting, and export in one place. Instead of starting the narrative from a blank page, I can start with a draft and focus on checking it.

Let the model write the first draft. Save your energy for the decisions that matter.