Day with AI: 8 Models, a Pile of Receipts, and the Answer to "Which AI Reads Your Invoices Best".

What I worked on

Tuesday. Quieter day — one big task instead of twenty small ones.

What I did

AI model benchmark for invoice extraction

Client project needed to pick the best AI model for reading invoices and receipts. So I took real documents and ran 8 different models on them — from premium tier to the cheapest ones.

What I tested:

• Extraction accuracy (tax IDs, amounts, VAT, line items)
• How they handle poor scans vs clean PDFs
• Speed and cost per request
• Edge cases — missing fields, non-standard formats

Result? Differences between models are surprisingly small on simple documents. But once you throw in a bad scan or non-standard layout, premium models win by a mile. Created a complete comparison report with HTML visualization.

Cold email pipeline

Morning session on the cold outreach pipeline. Tuning automation and flow.

Blog automation

Set up automatic "Day with AI" post generation — so at the end of each day, a blog post creates itself from source data (git log, daily log, sessions). Meta, I know.

Time: AI vs without AI

Task	Without AI	With AI

Benchmark 8 models on real data	3-4 days	~4h
Comparison report + HTML visualization	1 day	~30min
Cold email pipeline tuning	—	~1h
Blog automation	1 day	~1h
Total	~25 hours	~6.5h

What I learned

• Cheap models do fine on clean documents — premium is worth it only for messy stuff (bad scans, non-standard formats)
• Automated benchmark with HTML output is a great client deliverable — they see data, not just "trust me"
• Blog automation is a perfect example of AI saving time on things you just wouldn't do otherwise