Local LLMs in the Trenches: Building a Privacy-First TODO List Analyzer

As a vocal critic of large language models (LLMs) who nonetheless operates in an AI-saturated industry, I recently embarked on a practical experiment: Could a local LLM reliably categorize my daily TODO list items to track time allocation? The goal was simple—automate insights into my work patterns without compromising sensitive data. Here’s what worked, what failed spectacularly, and why hybrid systems might be the only viable path forward.

The Non-Negotiables: Privacy and Pragmatism

My requirements were strict:
- Local LLM only: As a manager, my TODO lists reference team members and confidential projects. Sending this to cloud-based AI providers like OpenAI was off-limits.
- Cron-able automation: The script had to run unattended, processing markdown files from my Obsidian vault.
- Zero format changes: My TODO structure (see below) couldn’t be altered—speed of entry trumped linting.
- Fixed categories: To combat LLM inconsistency, I predefined categories like call-meeting or incident response, forbidding ad-hoc inventions.
- Simple metrics: Output needed only counts per category, not reconstructed tasks.

# Daily Notes: 2025-06-15

### TODO
- [ ] Review Q3 roadmap

### Today
- [x] Debug API timeout (urgent)
- [ ] Morning Sync (09:00)

**Calls**
- [x] 1:1 with Alice

**Morning Routine**
- [x] Slack triage

Engineering the Pipeline

My TODO lists use hierarchical markdown, making them parsable via Python. Initial processing used simple conditionals:

if "**Calls**" in line:
    section = "calls"
elif line.startswith('- [x]'):
    category = categorise_item(line, section, ai, predefined_categories)

For AI integration, I deployed ollama with the 7B-parameter deepseek-r1 model—a compromise between performance and laptop-friendly resource use. The prompt enforced rigid rules:

Prompt Excerpt: "You MUST categorize items ONLY into: call-meeting, call-11, daily admin, incident response, PR work, documentation and planning, or other. NEVER invent new categories like 'security'."

Early tests seemed promising. For - [x] Swear at AI, it correctly returned "other". But reality soon intruded.

When AI Goes Rogue: Typos, Rebellion, and Workarounds

The LLM consistently defied instructions:
- Invented categories like security despite explicit prohibitions.
- Misspelled valid labels (e.g., documentaton instead of documentation).
- Added prefixes (*category*: call-11) or "corrected" my intentionally non-word categories (managering became managerial).

Reprimanding the model ("You’re a very naughty LLM!") in follow-up prompts changed little. So I built a hybrid approach:
1. Script-based substring matching for predictable items:

if "1:1" in item: return "call-11"
if "Slack" in item: return "daily admin"

   

This handled ~66% of items via rules, reducing AI exposure.

alt="Article illustration 2"
loading="lazy">

curl -d 'todo_list_completed,category=PR_work,host=workstation count=1' http://localhost:8428/api/v2/write?db=workload_stats

<img src="https://news.lavx.hu/api/uploads/local-llms-in-the-trenches-building-a-privacy-first-todo-list-analyzer_20250731_084521_image.jpg" 
     alt="Article illustration 1" 
     loading="lazy">

todo_list_items_count{db="workload_stats", category!="error"}

score = (1 - (retries / (max_retries + 1))) * 100  # 100% if no retries, 0% after max

<img src="https://news.lavx.hu/api/uploads/local-llms-in-the-trenches-building-a-privacy-first-todo-list-analyzer_20250731_084518_image.jpg" 
     alt="Article illustration 3" 
     loading="lazy">

The Bittersweet Verdict

Source: Ben Tasker