GenAI Cracow #31: JEV
Agenda
Opening Ceremony
Jev Sorting Email: Hits and Misses - by Pawel
Sorting a RAG Documents with Decision Models: Ollama for Ingest, OpenRouter for Queries - by Robert
Networking
Abstract
Jev Sorting Email: Hits and Misses
An insurer's letter and a payment got lost among ~200 unread threads a week, but handing everything to Claude Opus 5.5 is hiring a brilliant scientist to sort the office mail, at 95× Jev's input price. My Claude Code skill sends Jev only sender, subject and snippet, receives one of six categories and three yes/no flags as probabilities, not text, and lets Claude read the full thread only when Jev is unsure. Distrusting vendor benchmarks, I ran Jev against Nace.ai's Drex on 57 emails from my Gmail: both caught all 6 needing action, Jev's confident labels were 86% correct against Drex's 79% (median 281 vs 345 ms), and Drex cost less: $0.0017 versus $0.0022 for the batch. Identical Jev runs sent 19% and 26% to Claude, one filed an important FYI as Promo, and raising the 0.60 threshold to 0.70 gave 97% accuracy but sent 42% to Claude. Two thirds is live: raw JSON, the comparison, the threshold sweep.
Sorting a RAG Documents with Decision Models: Ollama for Ingest, OpenRouter for Queries
Looking for a general way to file knowledge-base documents for quick retrieval, I tested the approach of the open-source library Jevbox, where Jev picks the folder for every upload, on Polish legal acts: duplicate, amendment, excerpt, and where it belongs. Five model lines that return a probability per category instead of text ran through Ollama on an M1 Max laptop with 32 GB; for a chat assistant I am building for outside users, the same 83 test messages went to a local Nimble 9B (58 flagged correctly) and to Jev through OpenRouter (80), though their prompts differed. Filing a document when it arrives suits a local model: documents stay on the machine, a small model takes a few seconds per decision and answers identically, but it can cut a long text short without an error. Classifying while a user waits suits a hosted API: Clef (27B) on that laptop needed 36 seconds per document, but the text leaves the machine, OpenRouter’s EU endpoint offers no decision model, and the same request can return a different probability. Whether the Jevbox approach is the right way to manage knowledge stays an open question; the runs show that the category threshold has to be measured on the exact setup that will use it.
Speakers
Paweł Lipowczan
CTO and co-founder of 200IQ LABS, where he and Przemek Trybała design an agent system that does the company's office work. 17 years in IT: he started as a developer, then as Technical Lead at Automation House and Tigers chose technologies, designed no-code and low-code architectures (Airtable, Make, n8n) and led a team of analysts. He has spoken at Infoshare, NoCode & AI Event and BBConf4.IT. Outside work: sport.
LinkedIn
Rober Boguszewski Principal Architect at IG Group, an architect on its client technology portfolio and lead of the architecture team behind the division’s cloud migration. He assembled the division’s agent framework on an open-source base — repository structure, subagent definitions, standardised tool-server configurations — over a shared agent memory whose decision log the fleet reads before it acts. He chairs the division’s Architecture Review Board and turns its decision records into merge-blocking checks in CI. The division’s AI programme, which he runs, measures delivery cycle time and tracks adoption team by team. Away from work he builds AI products of his own, and runs the document-classification experiments behind this talk.
Partners
IG Group is a global fintech company that combines the scale and strength of an established FTSE-listed business with the innovation and entrepreneurial mindset of a technology-driven organisation. Operating through leading brands including IG, tastytrade, IG Prime, Freetrade and Independent Reserve, the Group has a presence in 18 countries across Europe, North America, Africa, Asia-Pacific and the Middle East. For more than four decades, IG Group has been helping people access financial opportunities through award-winning trading, investing and wealth-building products. Today, over 1.4 million customers worldwide use its platforms to access more than 19,000 financial markets, supported by innovative technology, data-driven solutions and a commitment to delivering exceptional client experiences.
Driven by a culture of curiosity, collaboration and continuous innovation, IG Group empowers its people to build products and services that shape the future of financial services while helping clients achieve their financial ambitions
CEE AI Hub