AI/ML

RETROSPEC

Machine-readable AI history knowledge graph — 170+ events, 11 epochs, designed for agents

Personal

4 technologies
2 key decisions
4 results

Problem

Problem

AI systems have no reliable machine-readable record of their own lineage. Static timelines are human-facing. Wikipedia articles are unstructured. When a future agent needs to trace why decoder-only transformers outcompeted encoder-decoder models, or understand what caused the two AI winters, there is no queryable, structured knowledge source for that. RETROSPEC was built to be exactly that — AI's own history textbook, structured for machines.

Approach

Approach

RETROSPEC organizes AI history across 11 epochs from philosophical roots (1300s) through the transformer revolution to the agentic era. Each event, entity, and relationship is stored as structured JSON with a consistent schema: causal links, influence graphs, lineage paths, and "why it stalled" post-mortems for failed approaches. The repository is designed to be consumed by LLMs and agents via a query API and an interactive Vercel web explorer. The influence graph lets any agent trace architecture decisions back to first principles.

Architecture

Architecture

RETROSPEC — system diagram

Historical SourcesJSON Schema (event…11 Epochs (1300s →…Influence GraphQuery APIVercel Web ExplorerLLM / Agent Consum…

Key Technical Decisions

Key Technical Decisions

Assembly Instructions — 2 Steps
01

Agents as primary audience

Human-readable timelines already exist. RETROSPEC's schema is designed so that jq queries, LLM tool calls, and agent retrievals return structured, self-describing results — not prose that needs further parsing. Every field has a stable key and consistent type across all 170+ entries.

02

Post-mortems for failed approaches

RETROSPEC documents why things failed, not just what succeeded. The Cyc project ($60M+, 600 person-years) is documented with a structured 'why it stalled' entry. An agent that studies these can calibrate its own confidence about approaches that have historically overpromised.

Results

Results

  • 170+ structured events across 11 epochs from 1300s to the agentic era
  • Influence graph enables lineage tracing from modern architectures back to first principles
  • Live web explorer deployed on Vercel
  • Machine-readable schema designed for LLM and agent consumption

Tech Stack

Tech Stack

PythonJSONVerceljq

Links