The Deployment Decade
The hard part of AI stopped being invention and became deployment. What that means for anyone running a transformation, and why being early stops paying.
Unfiltered takes on building AI systems, managing entropy, and leading product teams.
The hard part of AI stopped being invention and became deployment. What that means for anyone running a transformation, and why being early stops paying.
Eight years inside a bank and not one initiative failed on model quality. Root cause was the data contract and who was allowed to act on the output.
Notes from moving from enterprise software to physical AI. Retry stops being free, and dust beats accuracy.
Most enterprises are stuck in "Pilot Purgatory". They treat AI as a tech demo instead of a business capability. I've seen countless "innovation labs" burn millions on cool demos that never touch production data. The reason? A lack of "boring" infrastructure: data governance, CI/CD for models, and a clear path to ROI.
Read ArticleWhy Europe's banking sector is moving towards owned infrastructure. With regulatory pressure (GDPR, EU AI Act) mounting and data privacy becoming a board-level risk, relying solely on US-based hyperscalers is no longer a viable long-term strategy for critical infrastructure.
Read ArticleHow to filter the noise of new model releases and focus on what actually moves the needle. Every week there's a new SOTA model on benchmarks that don't matter to your business. This framework provides a scorecard for evaluating new GenAI capabilities against specific constraints: Innovation, Feasibility, and Viability.
Read ArticleAdapting high-velocity physics models for financial fraud detection. Traditional rule-based systems generate too many false positives, annoying customers and burying analysts. We took a different approach, borrowing Markov chain algorithms used in particle physics experiments to detect "interesting events" in noise.
Read ArticleLessons learned from building a Kafka architecture that handles 2 billion events daily. Scaling beyond "Hello World" is where the real pain begins. I discuss the operational nightmares of keeping a stateful cluster alive during peak loads, dealing with rebalancing storms, and the importance of idempotent consumers.
Read ArticleArchitecting low-latency inference systems before ChatGPT made it cool. In 2018, serving ML models in real-time (< 50ms) was a dark art. We explored the trade-offs between batch and real-time inference, eventually building a custom serving layer using C++ bindings for Python models.
Read ArticleExplorations into autonomous coding agents and their impact on developer velocity. I am currently running experiments to see if an AI agent can build a production-ready SaaS MVP in under 48 hours without human intervention. The results are... mixed, but promising.
Read ArticleTesting open weights models on consumer hardware. Can a Mac Studio replace a V100 instance for fine-tuning 7B parameter models? I'm benchmarking Llama-3 and Mistral on local silicon to see how far we can push privacy-preserving local AI.
Read Article