Notes from the foundry
Engineering essays on generative AI, retrieval systems, and what it takes to ship intelligent software to production.
Build Your Own Agent Workspace Before You Buy One
Buzz, Grok Bot, and every major suite are racing to own the room your agents work in. The shape they are converging on is copyable in about two weeks on infrastructure you already run, and doing that first tells you what you actually need.
Small Specialized Models (SLMs) vs Frontier Giants: The 10x Economics of Domain Distillation
Deploying a 400B frontier model for structured classification and routine data extraction is an economic blunder. Here is how enterprise teams train, quantize, and orchestrate 3B–8B parameter domain-specific models for 90% cost reduction.
Grok Build: Eight Subagents in Git Worktrees
xAI's terminal coding agent fans out to eight parallel subagents, each in its own git worktree. The isolation is the part worth copying.
Sub-500ms Voice Agents: WebRTC, Full-Duplex Audio, and Natural Turn-Taking Mechanics
Cascaded speech-to-text -> LLM -> text-to-speech pipelines produce awkward 2-second pauses that break conversational immersion. Here is how modern real-time voice architectures achieve natural turn-taking with WebRTC and speech-to-speech models.
How to Build Sub-Second Voice AI Agents With Twilio, WebSockets, and Small LLMs
Human conversation falls apart when latency exceeds eight hundred milliseconds. Here is the full engineering architecture for building real-time voice agents with bidirectional audio streaming, acoustic barge-in, and fast inference.
Fine-Tuning vs. RAG: An Executive Decision Guide for Engineering Leaders
Should your engineering team fine-tune an open-weights model or invest in a hybrid RAG pipeline? Here is the architectural and financial decision framework we use with enterprise CTOs to avoid six-figure engineering mistakes.
Grok Bot: Always-On Agents, Reviewed
xAI's always-on agent runtime, tested against real workloads: what the persistent execution model does well, and where it still bites.
Local-First AI & Edge Inference: Running Autonomous Agentic Workstations Privately
Cloud LLMs offer power, but data sovereignty, latency, and air-gapped security demand local inference. Here is how modern engineers build hybrid agent loops running quantized reasoning models locally on developer workstations.
Buzz Puts Your Agents in the Channel. Here Is What That Actually Changes.
Block shipped Buzz in July: an Apache-2.0, self-hostable workspace on Nostr where every participant, human or agent, is a keypair. It is pre-1.0 and genuinely interesting, and the interesting part is not that it looks like Slack.
Zero-Trust Agent Security: Non-Human Identities, Ephemeral Tokens, and Blast Radius Containment
Giving an autonomous agent a permanent API key with admin privileges is an invitation for catastrophic data exfiltration. Here is how to architect Zero-Trust Non-Human Identities (NHI), ephemeral scoped tokens, and deterministic safety proxies.
The 'RAG is Dead' Myth: Why Million-Token Context Windows Still Need Retrieval
Frontier models can read entire codebases in a single prompt, so why are enterprise search pipelines still failing? Here is why million-token context stuffing collapses on multi-hop queries, and how hybrid search keeps answers accurate.
Automating Legacy Enterprise Software With Browser Agents and Computer Use
Half of enterprise business value is trapped behind green-screen portals and ancient ERPs with no APIs. Here is how we build resilient browser automation agents with Playwright, visual grounding, and structured retry logic.
How to Build an Agent Team That Is Not Just One Agent Wearing Three Hats
Everyone is building agent teams now, and most of them are one prompt with role labels. Here is what actually separates a team of agents from an expensive way to call the same model repeatedly, and how to staff one.
The Shadow Agent Crisis: Managing Unsanctioned AI Bots Across Enterprise Infrastructure
Developers are quietly deploying personal CLI agents, Copilot extensions, and unapproved MCP servers connected to production databases. Here is the operational blueprint for discovering, cataloging, and governing shadow agents.
RAG Is Not Dead, It Is Being Replaced by Intelligent Context Routing
Dumping 50 chunked embeddings into a vector database was a 2024 shortcut. As reasoning models and million-token windows mature, modern architectures use hierarchical context routers, graph indices, and dynamic retrieval tiers.
AI Voice Agents Answer the Phone Now. What That Actually Does to Your Support Org.
Enterprise voice agents crossed from novelty to normal this year. The technology question is largely settled. The interesting questions are what happens to your escalation path, your metrics, and the people left holding the hard calls.
Cutting Enterprise LLM Spend With Context Caching and Token Budgeting
System prompts and enterprise schemas waste millions of tokens on every agent hop. Here is how we implement prompt caching, dynamic prefix alignment, and token budgets to drop monthly cloud AI bills by sixty percent.
Context Caching & KV Cache Tiering: How to Cut 80% Off Multi-Turn Agent Latency and Cost
Long context windows are cheap until you re-send millions of tokens on every tool step. Here is how prefix caching, prompt structure isolation, and hierarchical KV cache tiering turn an expensive loop into a performant system.
Shadow Agents: Somebody in Your Company Has Already Built One
Over a million agents have been built in Copilot Studio alone, most of them by people who are not engineers. Shadow AI is not a hypothetical governance risk, it is the current state of most enterprises. Here is how to find them and what to do that is not a ban.
What Agentic AI Actually Looks Like in Supply Chain and Manufacturing Operations
Supply chain is the function where agents have the clearest business case and the worst data. Here are the four workflows where AI agents earn their keep in manufacturing and logistics, and the two where they consistently disappoint.