Two ways to run keyes.ai services
The same engine — Vector, Memory, GitDB — runs as a managed service or inside your own cloud account. A short note on what each shape is for, and what's actually built.
Notes on GitDB, agent-to-agent workflows, and building a database for the AI-coding era.
The same engine — Vector, Memory, GitDB — runs as a managed service or inside your own cloud account. A short note on what each shape is for, and what's actually built.
AI-native search is a real category now — Exa alone is a $700M business — and the open technical question is the same one we keep running into. The retrieval layer either finds the right document or it doesn't, and at scale most stacks settle for "approximately".
AI customer service agents are a real category now — Decagon alone is a $4.5B business — and the most expensive failure mode in production is hallucinating a policy. A lot of that comes down to the retrieval layer.
AI is becoming a real participant in trading, and the memory layer has to satisfy two requirements that usually pull against each other — millisecond latency and 100% recall under audit. Here's how UQL approaches that.
Two memory services lead the agent-memory category — mem0 and Supermemory. Both are well-engineered for conversational personalization. Here's where ours sits and where each of theirs is the better choice.
Legal AI tools hallucinate 17–33% of the time per a peer-reviewed Stanford study. A lot of that comes from retrieval, not generation. Here's the memory-layer math, and what we benchmarked.
A look at real, documented cases where departing engineers walked out with the codebase — and what GitDB does to make that materially harder.
GitHub has been built for human developers for over 15 years. GitDB has been built for AI agents from day one. Both are correct designs — for different actors. A look at what each one fits.
The standard paths to 100% vector recall are GPU brute-force and mmap-backed indexes. Both work; both have a property that ruled them out for the workloads we kept seeing. A look at the design space and where we landed.
An honest look at the on-device data problem facing humanoid robots and self-driving stacks — and the embedded engine I built to fix it.
Three production vector databases dominate the category. A neutral look at what each one does, what we do, and how the trade-offs compare.
text-embedding-3-large produces 3,072-dim vectors, and most vector pipelines truncate them to stay performant. Here's what happened when we benchmarked the full dimension end-to-end.
A walk-through of the actual token math — where the savings come from, what they don't help with, and how I'd reproduce them in your codebase.
A walk-through of four coding agents shipping a feature in parallel — what each one did, where they handed off, and what the bill looked like.
The conventional wisdom on vector search is that you have to pick two of three — recall, speed, footprint. Here's what our engine does on the dbpedia-openai benchmark.
How GitDB coordinates multi-agent code workflows without ever putting source on a developer machine — and why a database, not a filesystem, is the right substrate for it.
AI coding agents spend 60–80% of their tokens just finding things in the codebase. That's grep hell. Here's what we built underneath instead.