Vinícius Raposo

Applied AI Reliability Engineer

AI systems you can leave on without babysitting.

Supporting project · Python

agenttrace

Tracing SDK, collector and dashboard that find the failing step across concurrent AI workflows, with cost attribution.

171 SDK + 124 collector + 41 dashboard tests; package and isolated-wheel checks; checksum-verified evidence. Provider answers, usage and prices in the demo are synthetic; live PostgreSQL/Redis is not verified.

Language: Python Updated: 2026-08-16 Visibility: public
View on GitHub Back to catalog

Without post-hoc run inspection, agent failures stay opaque—you cannot attribute cost, replay decisions step by step, or debug what the model and tools actually did on a failed run.

A run can be inspected after the fact.

Offline demo:

make demo-ui

Walkthrough

Deterministic replay of the issue-to-draft-PR run.
Deterministic replay of the issue-to-draft-PR run. Sample data.
Span timeline with token use and cost attribution.
Span timeline with token use and cost attribution. Sample data.

Why it matters

The default demonstration runs locally without credentials or hosted services. The absorbed issue/PR scenario proves path and branch refusals, test gating, explicit approval, draft-only intent, ordered audit events, and deterministic replay without contacting GitHub.

Verified proof

171 SDK + 124 collector + 41 dashboard tests; package and isolated-wheel checks; checksum-verified evidence. Provider answers, usage and prices in the demo are synthetic; live PostgreSQL/Redis is not verified.