Writing · IntelligenceOS
I publish the engineering work.
Notes on what I'm building, evaluating and getting wrong across production AI, financial systems, model adaptation and architecture.
Subscribe on Substack ↗
Latest · 90 seconds
The Model Is Not the AI System
The model is probably not the hardest part of your AI system. The hard part is everything around it: data, tools, state, permissions, evaluation, failure recovery and verification.
AI evaluationProduction agentsFinancial automationPayment systemsArchitectureRAGFine-tuningReliabilityFailure analysis
INTELLIGENCEOSYour First Fine-Tune: From Training Examples to a Working Python AppA beginner’s walkthrough with Unsloth: the dataset, the mistake, the evaluation, and the model running locally.2026-09-23Read ↗INTELLIGENCEOSThe Model Is Not the AI SystemWhat turns model capability into useful work: context, control, execution, and feedback.2026-09-22Read ↗INTELLIGENCEOSI Fine-Tuned a Model. The Failures Taught Me More Than the Winning ScoreWhat a classification win over GPT-4.1—and failures on unfamiliar regulatory questions—taught me about building useful domain intelligence.2026-09-11Read ↗INTELLIGENCEOSThe Model Is Not the AI SystemA model gives you capability. The harness around it turns that capability into intelligence that can actually work inside a business.2026-09-02Read ↗INTELLIGENCEOSThe Reframe: You’re Not Learning AIThe build is the curriculum. Choose a real friction, make one testable loop, and let failure write the syllabus.2026-08-21Read ↗INTELLIGENCEOSThe IOS Agent Layer — Building Reliable Multi-Step Analysis Pipelines for FinanceA production agent layer for document analysis is not a GPT wrapper with function calling. It’s a state machine with failure handling, a typed tool registry, deterministic routing…2026-06-19Read ↗INTELLIGENCEOSDomain Embedding Fine-Tuning — Adapting BGE to Finance and Risk DocumentsThis post addresses that gap. Not by switching models, but by teaching the one we have.2026-06-18Read ↗INTELLIGENCEOSGeneral Embeddings Are a Tax on Domain AIToday’s deep dive covers exactly this: how we built BGE-finance-v1, what the training data looks like, what Matryoshka fine-tuning gives you, and what a 24-point nDCG improvement…2026-06-08Read ↗