Writing
What I amworking out
Notes on data, automation, and systems that have to hold up inside real companies. I write when something took me effort to understand and I suspect it will take someone else the same.
- 6 October 20264 min read
I was about to compute customer balances with embeddings, and embeddings can't add
I'm putting a chatbot inside our ERP for shop-floor questions, and I started by embedding the tables so it could answer how much a customer owes. Today I understood that similarity search returns the k closest records and then sums whatever it returned: if the customer has 23 invoices and k is 5, the total is wrong and nobody notices.
- embeddings
- pgvector
- rag
- erp
- 21 September 20263 min read
The predictor tied with "the price always goes up", and the range theses were impossible from the start
I built a research system with agents that I could audit, added a predictor, and when tested against 8 years of data it tied with the dumbest rule possible: "the price always goes up". Then five range theses on Bitcoin fell on the same day, turned out to have been impossible from the start, and what ended up working were filters that tell the model "no".
- crewai
- agentes ia
- llm
- bitcoin
- 14 September 20263 min read
I capped the agentic loop at 5 iterations and it never got there; ARCA from a VPS still does not work
Two fronts in one week: an end-to-end project with Claude Code writing the code and me supervising, inside a loop of generate, review, audit and redo with a hand-set iteration cap. And automatic invoicing against ARCA, which got stuck not in the logic but in OpenSSL and in IP blocks against datacentres.
- claude-code
- agentes
- facturacion
- vps
- 7 September 20263 min read
I scaled scraping from 10 to 50 pages and Cloudflare stopped me
A RAG of agents that tracks the 3D printing market went from scraping 10 pages to 50, and that is when the 403s, the rate limits and Cloudflare showed up. How I got through it by using each site's own API where the wall was highest, and why the agentic loop turned out to be a separate problem.
- rag
- agentes
- web-scraping
- impresoras-3d