Patryk Szczygło
@patys_prog
I'm blogger, young programmer, multilingual man, interested in technology and travels. Lover of creating games and learning languages.
You might like
Chat summarisation experiment: - old: last 5 msgs, - new: summary from small model, issue: Omega msgs add noise, Testing: keep them for now, add prompt hint, watch results.
Weird discovery: Our AI gets noisy when past Omega messages are included in summaries. We'll try: - shrinking history to 20–30 msgs, - no filtering yet, just prompt it's Omega. If that fails -> filtering time.
Debated sales intent detection today. Keywords? Fast & cheap. LLM? Flexible & future-proof. We picked a small LLM with prompt-only context -> yes/no output. Cheap now, ready for more later.
Keywords work… until they don’t. We chose a tiny LLM for sales intent detection: - single prompt, - yes/no output, - minimal latency. Future-ready without burning budget.
Today's "overkill" is tomorrow's "bare minimum". That's why our sales intent detector is LLM-based (mini model). Prompt in, yes/no out. Cost-friendly, adaptable, and ready to grow.
Dependency resolvers don’t “just know” you’re fine. I had requires-python = ">=3.12.9". A package needed <4.0.0. Resolver: “Not compatible.” Solution: requires-python = ">=3.12.9,<4.0.0". Computers are literal. Humans are not.
TIL: requires-python = ">=3.12.9" ≠ >=3.12.9,<4.0.0. Dependency resolvers won’t assume you meant “less than 4” unless you say it. Future-proof your pyproject.toml.
Me: I have Python 3.12.9. Dependency: I work with >=3.8,<4. Resolver: “Not compatible.” Turns out you have to explicitly write <4.0.0. Machines don’t guess intentions, they check bounds.
Patch versions are supposed to be safe. Upgraded OpenAI SDK from 1.92.x -> 1.99.x via >= and suddenly ChatCompletion*ToolCall* types broke prod. Lesson: fast-moving SDKs? Pin exact versions.
Today’s bug: prod broke because OpenAI SDK “patch” release changed types 5 times in a week. Fix: lock to 1.99.1 until dust settles.
Working on better chat context. Swapping “last 5 messages” with a short summary from a small model. Problem: Omega’s past messages sometimes cause weird output. Current fix: limit to 20–30 msgs + tell model it’s part of Omega.
Today’s bug: - null where true/false should be, - “unhandled message” + “cannot translate typing union” errors, Cause? API drift + old libs. Fix? Upgrade, validate, and stop trusting yesterday’s behavior.
Something changed, our intention classifier started returning null instead of true/false. Lesson: if your logic assumes exact token boundaries — it’s already brittle. Treat prompts + tokenization like code: contracts, tests, versioning.
Thinking ahead: What if agents acted as internal project databases for dev teams? Memory + reasoning, scoped per project. Could reduce context loss — or just add noise. Anyone tried this?
New setup: - SalesAgent only at the start, only for sales topics, - Database, Primary, and Critic handle the rest, - Critic approves final message. - Flow is cleaner, scaling is easier.
We ditched AI Selector in Autogen. Too many misfires — SalesAgent was picked 50/50 with PrimaryAgent, even when it made no sense. Result: bloated, off-topic messages. We needed more control.
Our agent is finally getting memory. We're adding a database to store key insights, not just logs. It’s the first step towards actual continuity between conversations. #devdiary #AI #LLM
We're building file sync between Slack, Google Drive and our AI agent. Why? Because an agent that can't access your files isn't really helpful. Trust starts with not missing important context. #buildinpublic #aiagents
Dev diary #20 Right now we’re fully focused on getting three core features out the door: - file sync between Slack, Google Drive and the AI agent, - removing the selector function and replacing it with a graph-based approach, - adding a database so the AI can start saving memory
Dev diary #8 — We officially hit AI Agent bingo: everything that could go wrong, went wrong. Infinite loops, tool misuse, ignoring questions — the full set. Sometimes that’s how you learn.
United States Trends
- 1. Cynthia 59.2K posts
- 2. Good Thursday 29.3K posts
- 3. #GrabFoodMegaSalexหลิงออม 483K posts
- 4. #WorldKindnessDay 8,932 posts
- 5. Larry Brooks N/A
- 6. Rejoice in the Lord 2,364 posts
- 7. RIP Brooksie N/A
- 8. SUSDT N/A
- 9. #SwiftDay N/A
- 10. Happy Friday Eve N/A
- 11. RIP Larry N/A
- 12. PancakeSwap BNB Chain N/A
- 13. #thursdaymotivation 1,957 posts
- 14. #thursdayvibes 2,666 posts
- 15. Michael Burry 7,631 posts
- 16. Eddie Guerrero 3,945 posts
- 17. Namjoon 106K posts
- 18. Jeffrey Epstein 522K posts
- 19. Jesse Jackson 1,462 posts
- 20. Mikey 54.4K posts
Something went wrong.
Something went wrong.