Blog
How I Built Context Management That's Better Than the Frontier ProductsI built intelligent context compaction with persistent memory, adaptive RAG, and crash recovery — running on a self-hosted Qwen 3.6 on a Mac Mini. No cloud. No API costs. No data leaving the building.
The Table Saw ArgumentAI assistance is not a crutch — it is a force multiplier for a thoughtful engineer. Nobody asks a carpenter if they cut the wood by hand.
I Built an Enterprise AI Gateway From My PhoneNo laptop. No IDE. No terminal. Just Telegram messages to two AI agents. The abstraction layer between intent and implementation got thin enough that a phone keyboard is a sufficient development environment.
Your AI Doesn't Have Amnesia. You Just Haven't Built It a Hippocampus.LLMs can't form new memories after deployment. The brain solved this with a separate memory system. We built the same thing — externally — and beat fine-tuning on a compliance benchmark.
Building an Open Source AI Gateway: Full Visibility, Zero Vendor Lock-InCompanies adopting AI have zero visibility. Commercial gateways charge $2K-20K/month. I built the same thing with open source, running on a laptop, for $200/month.
I Trained a Domain Expert LLM That Ships Inside My ProductI needed a CMMC compliance model that runs on a laptop with no GPU and no cloud. Fine-tuning hit a ceiling at 84%. A stock 4B model with the right context at inference time hit 100%.
Why a Periodic Table of Data MattersWe organized 21 data modalities by how well AI models transfer between them — and discovered that data types cluster into six families, just like chemical elements. The gaps predict where undiscovered data types should exist.
Labeling Existing SharePoint Documents at Rest with File Extension MatchingThe free, undocumented path to retroactively labeling everything in SharePoint and OneDrive — and the Purview dashboard gotcha that makes everyone give up on it.
Calling Nellie a Memory Store Sold It ShortNellie isn't a memory store — it's hook-based context engineering middleware for Claude Code. And academic validation of the category just landed on arXiv.
I Cut My AI Bill in Half by Working FasterI was rationing a $200/month AI plan. Now I'm uncapped on $100. The bottleneck was never compute — it was workflow.
amp-rs: The Reference BoardIf AMP is the spec, amp-rs is the working board — the Rust reference implementation of Agent Memory Protocol. Clean, opinionated, and ready to run.
How I Built NellieEvery session started the same way — paste in context, burn half the window, start working. I was tired of this loop. So I built a memory system that remembers what matters.
Are We Solving Context Wrong?The AI industry keeps pushing bigger context windows. But context reset isn't a capacity problem — it's a signal problem.