I build AI infrastructure and agent orchestration systems.
Verdict is an autonomous development control plane. A goal goes in; a verified, independently reviewed, receipted change comes out.
Key capabilities:
- Goal-to-DAG orchestration: frontier planner decomposes work into a dependency graph; each node gets its own model and worktree (demo run proof)
- Admission control: live eligibility ladder (
DISCOVERED → ENTITLED → HEALTHY → AVAILABLE → TASK_ELIGIBLE → SELECTED) keeps cheaper capacity first; every dropped candidate carries a named reason (source) - Same-node reroute: quota, rate-limit, timeout failures cool down the route or provider; the same node reassigns to another admitted model until pool exhaustion (recovery intelligence)
- Independent review: reviewer excluded from all implementers; route-level or family-level independence enforced (review policy)
- Tamper-evident receipts: SHA-256 digest of the event log; any later edit is detected on verification (receipt module)
Proof write-up: VERDICT_PROOF_CASE_STUDY.md.
Credential-free demo (no API key, no gateway, no network):
pip install verdict-core
verdict quickstart --non-interactive --dry-runThe demo makes one deterministic routing decision and names every excluded candidate. Watch the terminal recording or inspect the committed run proof.
Read more: main README, orchestration architecture, claims audit.
| Repo | Purpose |
|---|---|
| verdict-node | TypeScript envelope enforcement and task classification for Node.js; published as @bodanglin/verdict-node 0.2.0 |
| verdict-cockpit | Source-available fixture-mode viewer (not published; active development) |
Python, TypeScript, FastAPI, async/event-driven systems, policy enforcement, admission control, deterministic recovery, reproducible testing, and evidence-backed technical communication.