Video summary

Excel-таблицы с Goose AI и локальной LLM | Пишем инструмент для аналитики

Main summary

Key takeaways

Technology

Tech goal / problem addressed

  • Build an analytics tool that uses a local LLM via an “agent” framework (Goose AI) to clean and analyze Excel reports.
  • Typical input data issues in large Excel files (thousands–tens of thousands of rows):
    • Broken/invalid rows (e.g., negative amounts that shouldn’t exist)
    • Unknown or disallowed goods/products
    • Outliers in numeric amounts

Why an LLM alone isn’t enough

  • Plain LLM usage is risky because LLMs are generative/predictive, not deterministic.
  • The agent helps by having the model produce code that processes the dataset in a fully deterministic and repeatable way (same input → same output, daily reruns).

Pipeline / workflow (as demonstrated)

  • Raw data: XLS/XLSX file (downloaded daily)
  • Local LLM agent configured with Goose AI Framework
  • Outputs:
    • Cleaned data
    • A short report
    • A separate sheet/logging of errors and reasons + recommendations

Implementation steps & features shown

Install & configure Goose AI

  • Set up a local provider for the model.

Load local model

  • Mentions using “Quent 3 encoder” and checking context/settings.

Generate test data

  • Python script creates a synthetic Excel dataset:
    • dates, names/products, amounts
  • Test anomalies include:
    • “unicorn product”
    • negative amounts

Run the agent (developer option + task prompt)

  • The agent calls tools, writes code, runs it, and performs self-correction when errors occur.

Debugging a template/provider incompatibility

  • Encountered a prompt/template error due to incompatibility between the Goose/Ginger model template requirements.
  • Fixed by editing/replacing the template (no reboot needed).
  • After correction, the agent successfully produced results.

Result format (what the report contained)

  • Excel output with multiple sheets:
    1. Errors sheet: reasons for each issue + what it is + recommendations from the model/agent.
    2. Cleaned data sheet: cleaned dataset with invalid outliers removed/fixed.
  • The agent successfully cleaned raw input data and generated a report.

Key design claims

  • Privacy / security: using only local model + local tools so confidential data doesn’t leave the machine.
  • Determinism: the agent executes code-based processing, not free-form generation.
  • Extensibility: not limited to Excel—agents could use tools for:
    • file system access
    • Excel/PDF documents
    • databases
    • search engines API and other integrations

Additional tutorial: “vibe coding” a local web app

  • Builds a web application so non-developers can upload/process Excel files via a UI.
  • Approach:
    • Used cldcode (coded by an AI coding workflow) with frontier mode and a model mentioned as “45”.
    • Rationale: local models “still write code mediocrely,” so a different coding model was used.
  • Iterative debugging:
    • First run produced a white page → logs/screenshots sent to fix.
    • Later iterations got processing working; eventually the app produced results (cleaned data + report), though templates differed slightly from the console version.

Main speakers / sources

  • Speaker: “M.” (as credited near the end of the subtitles)
  • Tool/framework sources:
    • Goose AI Framework (agent for local LLM tool use and deterministic code execution)
    • cldcode (used for the web app scaffolding via “vibe coding”)

Original video