Insights
Research from the production floor.
Field notes, research and published writing on the unglamorous work of putting AI into production — evaluations, retrieval, structured outputs and the gap between a demo and a system.
Executive research
7 enterprise AI opportunities: insights from 20+ C-level executives
Patterns from conversations with more than twenty C-level leaders: where enterprises are deploying voice AI, autonomous agents, self-hosted models and AI training today.
Compliance
AI-driven data classification for security and compliance
Why classification is the foundation of data security, and how language models make it continuous, contextual and auditable across GDPR, CCPA and HIPAA obligations.
Published with
Technical writing in the wild
Work we have published with Modal, Timescale and Instructor — the same retrieval, embedding and schema patterns we put into client systems.
- ModalBeating proprietary models with a quick fine-tuneRunning a parallel hyper-parameter grid search across 50+ GPUs on Modal to fine-tune an open embedding model that outperforms OpenAI’s text-embedding-3-small.
- ModalEmbedding English Wikipedia in under 15 minutesUsing serverless cloud GPUs to embed the whole of English Wikipedia, a demonstration of how far embedding workloads can scale with a few lines of code.
- TimescaleRAG is more than just vector searchWhy production RAG needs SQL-grade metadata filters alongside semantic search, and how combining lexical and vector retrieval in one database simplifies the stack.
- InstructorBad schemas could break your LLM structured outputsBenchmarks on GSM8k with GPT-4o and Claude showing how schema design and JSON mode versus structured outputs change accuracy, with practical schema-engineering guidance.
- InstructorSmarter summaries with fine-tuning GPT-3.5 and Chain of DensityAn early demonstration that fine-tuning can deliver large gains on focused tasks. The piece reached #3 on Hacker News.
- InstructorRAG is more than just embedding searchThe limits of embedding-only retrieval, and why scaling beyond a demo requires query understanding and careful metadata filtering.
Want this thinking applied to your stack?
Book a 30-minute consultation and leave with a concrete automation roadmap.