Multi-agent legacy documentation
An orchestrator, a deterministic analyzer, parallel explorer agents and an independent reviewer turn an undocumented codebase into a verified, 16-section technical report.
- Result
- 16 sections
every claim verified against cited lines - Role
- Architect and builder
- Domain
- Client demo, embedded systems
- Period
- Aug 2026 – Oct 2026
Problem
Legacy firmware and enterprise code often has no living documentation, and LLM summaries hallucinate. Modernization starts with understanding, and understanding needs to be trustworthy.
Approach
Split the job between deterministic tools and agents, and never let an agent's claim through unchecked.
What I built
- A master orchestrator plus a static analyzer for call graphs, interrupt handlers and registers.
- A doc lead that plans workstreams, and parallel explorer sub-agents with read-only tools.
- An independent reviewer that checks every claim against cited source lines (verified, corrected, or flagged for an expert).
- A visualizer that generates Mermaid diagrams and repairs syntax errors automatically, plus an assembler that produces a 16-section report with a traceability matrix.
- Replay of recorded runs for reliable demos, ingestion by zip, folder or repo, and switchable model providers.
Outcome
- Documented a 32-file, 3,000-line firmware sample (200+ magic numbers, cyclomatic complexity up to 51) in a 15-minute live demo.
- The reviewer stage turned a plausible draft into one an engineer can trust.