Aherontas

How I Built a Signed Audit Trail for an LLM System (And Broke It With One SQL Update)

Logs, traces, governance, compliance, audit trails... those are some of the words that come up every time we discuss putting LLM features and AI agents into production.

Most systems already have logs. The real question is different:

Can you prove that the logs you are showing today are the same records the system wrote six months ago?

Part 1/7 of my PyCon Greece 2026 talk, From Prompt to Proof: a Presidio guardrail masked the ΑΦΜ(tax number in Greece) label but leaked the actual tax identifier to the LLM while returning HTTP 200.

Unfortunately we will see this kind of cases more and more with AI rise. I don't believe it is the only app that could do relevant labeled searching in faces etc.

A same to not have them anymore :P

Curious how you handle feed evolution over time. When an RSS source changes structure (fields added/removed, summaries truncated, etc.), do you normalize to a fixed schema or store the raw payload alongside a best-effort normalized version? Longitudinal datasets tend to get tricky there.

I am doing exactly the same, also writing down some raw ideas I have in it to not forget them.

That's correct, the slides are mostly showing how MCP - A2A work and basic architectures between them and Agents in general. I will upload probably today the presentation slides in my blog.

It just builds a ChatGPT like UI that has Agents to serve you either latest news on a specific tech topic (via fetching them via MCP Servers for: Brave Browser, Hackernews and Github for trending repos or repos fetched in the response). It also supports giving you back general knowledge answers. The whole purpose of the repo is to be a general example of how Agents speak with Agents and how to pair them with MCP servers and why we need MCP Servers(for example better run all calculations not in LLM because they hallucinate, but via code and then give the answer back to the LLM). I hope it clarifies what it does now.

Haven't tried balm, I will give it a shot, I am really curious to see all the features it supports.