Back to work

AI Infra + Governance

Fleetrac

AI governance control plane

Building an AI governance platform that ingests OpenTelemetry traces, detects drift and policy violations across simulated agent systems, and drives approval-gated remediation through a FastAPI/SQLAlchemy backend and Next.js operator console.

Role
Independent Developer
Timeframe
2026 – Present
Status
In progress
Links

Product screenshots

Context and problem

Agentic systems need governance beyond static guardrails: runtime observability, incident correlation, accountable ownership, and approval-gated remediation when drift or policy violations emerge.

Challenge

Give operators a vendor-neutral way to detect, correlate, and remediate risky agent behavior from live telemetry—not just log dashboards.

Constraints

  • Independent project—simulated agent systems rather than production fleet
  • Must normalize heterogeneous telemetry from multiple LLM provider patterns

Key contributions

  • Built an AI governance platform that ingests OpenTelemetry traces to detect behavioral drift, prompt-injection risk, and tool-scope violations across 10 simulated agent systems.
  • Developed a FastAPI/SQLAlchemy remediation workflow that correlates risk events, escalates incidents in rolling 15-minute windows, and supports approval-gated policy enforcement and post-action verification.

Architecture

Key engineering decisions

  • Used OpenTelemetry as the vendor-neutral ingestion layer across simulated agent providers
  • Built event-driven correlation with dynamic risk windows instead of single-span alerting
  • Connected detection to approval-gated remediation rather than autonomous auto-fix

Security and reliability

  • Deterministic severity progression from parent-child span attributes
  • Post-action verification step after approval-gated remediation

Outcome

Delivered a working governance platform across 10 simulated agent systems with OpenTelemetry ingestion, rolling risk-window correlation, and approval-gated remediation workflows.

Limitations

  • Simulated agent systems—not yet deployed against a production agent fleet

Next steps

  • Expand provider adapters beyond current OpenTelemetry ingestion paths
  • Add policy-as-code enforcement hooks for remediation actions

Tech stack

  • FastAPI
  • Next.js
  • SQLite
  • OpenTelemetry
  • SQLAlchemy