Argus by Perciqa

The Problem

AI agents fail silently. A customer service agent hallucinating a refund policy, a research agent fabricating a citation: both return 200 OK. Traditional APM tools like Datadog and New Relic track system health, not agent reasoning. Nobody can see what the agent actually decided, what it cost, or whether the output was any good.

Teams shipping agents are flying blind. Debugging is guesswork, cost tracking stops at the GPU, and quality is unmeasured until users complain. The AMD hackathon gave me a constraint that sharpened the design: build the whole thing AMD-first, with zero CUDA dependencies, proving that serious AI infrastructure can run on non-NVIDIA hardware.

The Solution

Argus is a single tool that closes the reliability gap in three moves: trace every decision, price every inference, and grade every run.

• 1. Trajectory tracing: capture the full decision tree of every agent run, every reasoning step, tool call, and model selection, as a replayable structured trace with an interactive viewer.
• 2. Inference FinOps: track cost per task, not per GPU, with local versus cloud breakdowns. Local inference on AMD hardware counts as $0.00, surfacing the real savings of self-hosted inference.
• 3. Eval in production: every live trace is scored by an LLM-as-judge eval engine running DeepSeek V4 Flash via Fireworks AI serverless, with a local Ollama fallback when no API key is set.

Results

Architecture

A Python SDK intercepts OpenAI-compatible calls with zero extra code and streams traces over HTTP to a FastAPI server backed by SQLite. The server runs the eval engine and pushes live updates to a Next.js dashboard over WebSocket.

The eval engine is AMD-first by design: no CUDA dependencies, a ROCm-compatible architecture, and a FinOps layer that treats local GPU inference as free. Instrument an agent with @argus.trace, and every call is traced, costed, and scored automatically.

Recognition

Awarded for outstanding performance in the AMD Developer Hackathon: ACT II (July 6 to 13, 2026), for successfully completing and submitting a solution based on AMD Developer Cloud, AMD ROCm, and Gemma. Issued by lablab.ai, part of NativelyAI.

Download certificate (PDF) View the hackathon