Blog

Privacy-First AI: Why We Built an AI Assistant That Never Calls Home

31 May 2026

How Terraphim AI achieves privacy-first AI by moving knowledge graphs instead of data, using deterministic automata, embedded databases, and WASM deployment.

Terraphim RLM Skills: Local-First Recursive Language Model Orchestration

16 May 2026

How we shipped four native RLM skills for local-first deterministic orchestration with capability-based routing and persistent knowledge graphs.

The Latency Problem Nobody in AI Wants to Discuss

15 May 2026

Real-world API latency benchmarks showing Perplexity at 3,188ms, AWS Bedrock at 1,086ms, and Companies House at 121ms. Why the AI industry normalised slow APIs and what engineering discipline looks like.

Learning via Negativa: How Terraphim Remembers What You Keep Getting Wrong

14 May 2026

How Terraphim's Learning via Negativa system captures failed commands, builds a correction knowledge graph, and prevents AI agents from repeating the same mistakes.

Sentrux Quality Gates: Closing the Feedback Loop on AI-Generated Code

13 May 2026

How Sentrux structural quality sensors integrate with the AI Dark Factory to measure modularity, acyclicity, and complexity before and after AI-generated code changes.

zvec vs Terraphim: Two Paths to Semantic Search

12 May 2026

A deep dive comparing Alibaba's zvec neural vector database with Terraphim's graph-based approach to semantic search -- when to use which, and why hybrid retrieval is the future.

Terraphim GitHub Runner: AI-Powered CI/CD with Firecracker MicroVMs

11 May 2026

How Terraphim GitHub Runner combines LLM-powered workflow understanding with Firecracker microVM isolation for secure, private, and fast GitHub Actions execution.

The Context Graph: Why RAG Is Not Enough

10 May 2026

Why Retrieval-Augmented Generation is insufficient for production AI, how GraphRAG and ontologies extend it, and where context graphs fit in the progression from text chunks to structured knowledge.

Rust Binary Optimisation for Minimal Docker Images

9 May 2026

Practical guide to optimising Rust binaries for Docker: musl static linking, link-time optimisation, codegen-units, and scratch containers for sub-15MB images.

Firecracker VM Snapshots: Layerfile-Performance Build Systems in Rust

8 May 2026

Architecture and implementation plan for a Firecracker-based build system with automatic VM snapshots, file dependency tracking, and intelligent cache invalidation. The Layerfile approach rebuilt in Rust.

Dangerous Commands Guard: Why Prompt-Level Security Is Not Enough for AI Agents

17 April 2026

Why prompt-level security fails for autonomous AI agents, and how a pre-execution hook with deny-list patterns catches destructive commands before they run.

Deploying an AI Dark Factory: 10 Bugs in 48 Hours and What They Taught Us

17 April 2026

A field report from deploying autonomous AI agent factories that pick up Gitea issues, write code, submit pull requests, and close bugs without human intervention.

The Domain Model as Quality Thread: How JSON Schema Survives From Whiteboard to Production

17 April 2026

How a single JSON Schema-validated domain model carries quality from inception through to production operations, with 18 artefacts and 5 gate reviews.

We Tried Three Task Coordination Tools Before Realizing Gitea Already Had the Answer

17 April 2026

How we replaced three bespoke coordination systems with Gitea's Robot API and PageRank-based issue prioritisation for AI agent workflows.

Why We Chose Gitea Over GitHub for AI Agent Coordination

17 April 2026

Why we moved our AI agent coordination layer from GitHub to self-hosted Gitea: rate limits, PageRank prioritisation, privacy, and package registries.

GPU-Accelerated Knowledge Graphs vs Deterministic Automata: What I Learned Building Both

17 April 2026

Building both GPU-accelerated knowledge graphs and deterministic Aho-Corasick automata taught me that the technology choice matters less than curation discipline.

The LLM Proxy Nobody Asked For (And Why We Built It Anyway)

17 April 2026

Building a Rust LLM proxy with Aho-Corasick keyword routing, multi-provider fallback chains, and deliberate failure injection to prove it works.

What the METR Study Actually Proves About AI-Assisted Engineering

17 April 2026

The METR randomised control trial found developers 19% slower with AI tools. Why the J-curve is real and frontier teams redesign everything, not just their tools.

The Retrieval Cost Spectrum: Why Your AI Agent's Memory Architecture Matters More Than Its Model

17 April 2026

Three approaches to AI agent retrieval -- traditional RAG, agentic RAG, and graph embeddings via Aho-Corasick -- compared by cost, latency, and determinism.

Teaching AI Agents to Learn from Their Mistakes: From FINAL Bench to Production Self-Correction

17 April 2026

How the FINAL Bench declarative-procedural gap inspired Terraphim's learning capture system: persistent failure memory, Aho-Corasick query expansion, and the MetaCog self-correction scaffold.