• 

  ROOTRACE 2.0

ROOTRACE 2.0 — The AI That Must Prove Its Fix

AI proposes. Evidence decides.

Inspiration

AI coding assistants can suggest ways to optimize slow applications in seconds. But a suggestion is not proof.

A query might become faster while returning incorrect results. A plausible index might produce no measurable improvement. A benchmark might look impressive because of measurement noise rather than a genuine optimization.

We built ROOTRACE to answer a critical question:

What if every AI-generated performance recommendation had to prove itself before being accepted?

ROOTRACE combines NVIDIA Nemotron-powered hypothesis generation with controlled experiments, independent correctness checks and reproducible evidence.

Our goal is to move from AI-assisted guessing to evidence-driven engineering.

What it does

ROOTRACE is an AI-assisted performance investigation and verification system.

It transforms a performance problem into a structured investigation:

Performance problem → AI hypothesis → Controlled experiment → Correctness verification → Evidence-backed verdict

Instead of trusting an AI recommendation, ROOTRACE evaluates competing hypotheses and distinguishes verified improvements from inconclusive or untested suggestions.

The system includes:

• AI hypothesis generation: NVIDIA Nemotron provides candidate explanations and optimization ideas.
• Controlled performance experiments: Candidate fixes are measured against a baseline.
• Correctness verification: Performance improvements must preserve expected behavior.
• Independent evaluation: Deterministic rules assess experimental results.
• Evidence-backed verdicts: Outcomes are supported by recorded measurements.
• Performance Courtroom: An interactive interface presents hypotheses, results and supporting evidence.
• Verification tools: Published SHA-256 hashes allow independent verification of evidence files.

The public judges' demo uses recorded evidence and static pages. It does not make live model calls or expose the backend.

How we built it

ROOTRACE combines an AI reasoning layer, a performance experimentation layer, a deterministic evaluation layer and a browser-based evidence presentation.

Technology

NVIDIA Nemotron + Nebius Token Factory

We integrated NVIDIA Nemotron through Nebius Token Factory for AI-assisted performance hypothesis generation.

The recorded model interaction is preserved separately from the locally measured performance results.

Python, FastAPI and DuckDB

Python powers the experimental workflow and evaluation utilities. FastAPI supports the backend architecture, while DuckDB provides the database workloads used for performance investigations.

Independent correctness and evidence verification

Experiments capture raw trial measurements, evaluate competing hypotheses and preserve evidence artifacts.

The verification workflow supports independent SHA-256 checks and

Built With

• duckdb
• fastapi
• javascript
• nebius-token-factory
• next.js
• node.js
• nvidia-nemotron
• pytest
• python
• react
• typescript
• vercel