Software Development & Infrastructure Insights
Your people already use AI assistants, and the assistants know nothing about your company. The usual fix hands the model everything a search returns. PremAgentic decides what the model is allowed to know, before anything is retrieved. Here is why we built it, and what we keep seeing inside companies.
Read more →
Ask an AI assistant for a recommendation and you get one answer with a handful of sources. This post ranks 22 tactics for becoming one of those sources by the evidence behind each, from letting AI crawlers in to adding an llms.txt file, with a plain verdict and a confidence rating for every one.
Read more →
Add an approximate index to a permission-filtered vector search in PostgreSQL, and the people allowed to read the least start getting the fewest answers. Nothing errors and nothing leaks. The results just come back short. Here is the mechanism, measured, and the two fixes.
Read more →
LTFS is one specification with one reference implementation, which makes it sound like a solved cross-platform problem. It is not. On Windows it is a command suite that assigns a drive letter. On Linux and macOS it is a FUSE filesystem mounted at a path. Here is what actually differs, including a macOS build defect that costs a day if you have not seen it before.
Read more →
We ship two Model Context Protocol servers. One runs inside a desktop application on the user’s own machine. The other runs as a hosted service behind an account. They speak the same protocol and share almost none of the same decisions. Here is what differed, and why the deployment model decided nearly all of it.
Read more →
We run two MCP servers with opposite security postures. One can delete things. The other cannot change a single byte. Both are correct, because the question is never "is this secure" in the abstract. It is who absorbs the damage when the assistant gets it wrong.
Read more →
Most small-business websites are written to serve anywhere, which is why they win nowhere. We just shipped a site for a mobile coffee bike in Sierra Vista, Arizona, and it commits to its county on every page. Here is why tying a brand to a place is the strongest move a local business can make.
Read more →
Thousands of MCP servers shipped in the last year, and most are thin wrappers with every permission turned on. We built the one inside Hiberden, an LTO tape archiving app, the other way. Here is what it gets right about tool design, permission tiers, and packaging, and what anyone building an MCP integration should steal.
Read more →
A Karpathy-style Obsidian vault is a great memory for one developer and one AI session. It starts to break down when a whole team of humans and agents depends on it. A local pgvector index in front of the vault is the fix: here is the token math, the shared-memory payoff, and the honest costs.
Read more →
Most companies experimenting with AI do not have a model problem. They have a knowledge problem. Here is how we are building a governed knowledge layer that connects company information to Codex, Claude, ChatGPT, and future AI tools without surrendering control of the data.
Read more →
Almost every pgvector tutorial teaches semantic search over text and stops there. But pgvector is a nearest-neighbor engine, not an embeddings feature: for structured data where you already know what makes two things alike, a hand-built feature vector you can normalize, weight, and explain beats a black-box embedding. Here is the distinction, with a worked example from a real "find similar players" build.
Read more →
We ported a production .NET application to Rust to find out where the language actually earns its cost. The honest answer surprised us: a rewrite is almost never about which language is better. Here is the scorecard, the hidden costs nobody quotes, and the four questions that decide it.
Read more →
The web is filling with AI-generated slop, and the models your buyers use are being tuned to ignore it. Here is why getting cited by AI now matters more than ranking on Google, and what executives should do about it.
Read more →
inotify doesn't work on network-attached storage. Polling a 500,000-file archive takes two minutes per cycle. ZFS's copy-on-write block pointer tree was a filesystem CDC system all along, making change detection O(changes) instead of O(total files).
Read more →
We replaced a 765-line multi-engine OCR pipeline with a single 0.9-billion-parameter vision model running on a Mac. The new system is simpler, more accurate, and every byte of data stays on premises.
Read more →
Generative AI will transform how production companies use their archived footage. But AI cannot work with unstructured video files. The companies that index their archives now will have a massive head start.
Read more →
A client was spending thousands per month running every AI coding task through premium models like GPT-4o and Claude Opus. We restructured their workflow to use the right model for the right job and cut their bill by 80%.
Read more →
LTO tape stores data at about half a cent per gigabyte, and an unmounted cartridge is ransomware-proof by design. A petabyte-scale operator walks through how to size it, plan retrieval, and avoid the expensive mistakes before you buy.
Read more →
AI tools can build a website in minutes. That speed comes with real risks: security holes nobody audits, over-engineered code nobody understands, and a web that is starting to all look the same.
Read more →