TRACEBACK-Powered-by-Shadow-engine

Landing page
Workflow running
Verified report

Inspiration

Automated decisions often give people an outcome without explaining what was measured, what could change, or what remains unknown. A chatbot may generate a plausible but unsupported explanation. We built TRACEBACK to replace guesswork with a verified investigation.

What it does

TRACEBACK investigates synthetic loan decisions. After one click, a Google ADK agent autonomously:

loads and reproduces the decision;
calculates financial factors;
probes the model with controlled changes;
searches for bounded recourse;
runs ACTUAL-versus-SHADOW simulations;
compares similar synthetic cases; and
produces a verified, downloadable Case File.

Every statement is classified as a confirmed fact, possible reason, simulated outcome, or unknown. TRACEBACK never claims to reveal hidden causality or guarantee approval.

How we built it

TRACEBACK uses:

Gemini 3.7 Flash through Vertex AI;
Google Agent Development Kit;
Cloud Run and Firestore;
Streamlit, Python, and Pydantic;
scikit-learn for the synthetic decision model; and
deterministic tools for calculations, model reruns, recourse, and verification.

Gemini coordinates the workflow and writes a qualitative summary. Deterministic Python independently reconstructs every authoritative result and runs nine verification checks before displaying or storing the report.

Challenges

Our main challenge was separating agent autonomy from factual authority. Gemini is useful for coordinating tools, but calculations and model results must remain reproducible. We therefore allow the agent to sequence bounded tools while Python owns the numbers and final verification.

We also designed recourse carefully. TRACEBACK tests a fixed set of changes against the same synthetic model instead of generating unrestricted financial advice.

What we learned

Trustworthy agents need clear boundaries. Language models are effective orchestrators, while deterministic tools are better authorities for calculations and verification. We also learned that counterfactual results are valuable only when uncertainty and limitations are clearly communicated.

Accomplishments

TRACEBACK completes a multi-step workflow after one action, passes 29 automated tests including a live Gemini/ADK test, stores verified Firestore audit records, and produces a portable JSON Case File with provenance and integrity metadata.