Live
Plugin
A standalone memory runtime for AI agents that separates retrieval, historical recall, temporal validity, current work, trust, decisive selection, traceability, unfinished work, and explicit memory writes from permission to influence present behaviour.
Problem
Long-running AI agents accumulate files, conversations, decisions, failures, instructions and changing project state, but ordinary retrieval cannot tell historical evidence from current guidance, relevant material from authoritative material, or remembered information from permission to act on it.
Outcome
A standalone remembering runtime where canonical project history is retrieved strongly, routed by purpose, interpreted temporally, conditioned on the work being done, gated by standing, reduced to decisive provenance-bearing evidence, recorded in an immutable trace, with unfinished work tracked and explicit memory actions accepted without granting the writer authority over behaviour.
Book: Memory From First Principles
Capstone: The Memory From First Principles application: a real coding-agent plugin implementing the book’s earned path from canonical project history through retrieval, routing, temporal state, safe framing, trust, decisive selection, durable traces, open loops, and explicit memory actions.
Live
System
Evidence-first AI systems where claims, decisions, policy boundaries, failure states and replay remain explicit.
Problem
AI systems can cross from model output into decisions and actions without preserving the evidence, claim boundaries and policy state needed to audit or replay them.
Outcome
A bounded decision system whose important claims, evidence, decisions, failures and replay state remain explicit across runtime boundaries.
Live
Runtime
A relation-aware embedding runtime that measures identity, compatibility, preservation, and policy-scoped usability instead of trusting raw vector similarity.
Problem
Embedding systems often treat proximity as permission: if two vectors are close, the application behaves as though meaning, equivalence, compatibility, or preservation has been established.
Outcome
A measured runtime that records space identity, evaluates geometry and transformations, preserves evidence, and grants usability only for scopes that passed declared tests.
Book: Embeddings From First Principles
Capstone: RELATE 1.0: the embedding runtime, corpus, benchmarks, preservation profiles, and replayable evidence behind the book.
Live
Runtime
A durable runtime for putting model intelligence inside an inspectable process with explicit state, authority, effects, verification, acceptance, replay, and deterministic next-operation policy.
Problem
A model call can produce useful intelligence without producing a reliable process: chat history hides state, successful requests are confused with completed work, model claims become evidence, retries repeat side effects, and the system has no durable answer to what should happen next.
Outcome
A runtime where stochastic cognition is bounded by deterministic state projection, explicit authority, preserved observations, verification, acceptance, replay, and a recorded policy for choosing the next operation.
Book: Applied AI
Capstone: The Applied AI capstone: one task carried from directive and context through model cognition, controlled effect, bound verification, acceptance, replay, and an audit of the joints between them.
Live
System
Practical techniques and interactive tools for using AI as a thinking, exploration and creative partner rather than a one-shot answer machine.
Problem
Most AI interfaces encourage one-shot prompting even when the real task is exploration, reframing, comparison, synthesis or creative development.
Outcome
A reusable set of human + AI thinking techniques that can be applied as structured workflows instead of improvised prompts.
Book: Freestyle Cognition
Capstone: Turn cognition techniques into reusable interactive workflows and tools.
Reference
System
A DSPy-backed repository-repair loop that turns language-model program changes into candidates evaluated by independent evidence before promotion.
Problem
Language-model programs can optimize toward convenient feedback while confusing a better score, a valid patch and a trustworthy production change.
Outcome
A controlled repair loop where the LM can diagnose and propose, but ordinary software applies patches, checks scope, withholds independent evidence and records a separate promotion decision.
Book: DSPy From First Principles
Capstone: Chapter 20 repository-repair capstone: repair-run-20260901-013940.
Building
Tool
A user-owned browser AI layer that can filter, amplify, search, organize and act on information while keeping context, provenance and authority inspectable.
Problem
The browser contains the user’s real information environment, but today context is fragmented across pages, tabs, messages and services while AI assistance is usually trapped inside isolated provider interfaces.
Outcome
A user-controlled browser layer that can find, filter, amplify, retain and act on information using explicit policies, typed tools and inspectable evidence.
Book: Browser AI From First Principles
Capstone: A user-owned browser AI coordinator and Digital Lens built on the Browser AI Observatory, typed tools and explicit authority boundaries.