Skip to content

Failure Forensics

An AI reliability platform that finds where a bad AI answer first went wrong, shows how the error spread, and proves it with evidence.

Project at a glance

Item Detail
Organisation Undergraduate Artificial Intelligence Club, University of Alberta
Build window 2 Nov 2026 → 26 Feb 2027 (8 sprints)
Final demo First or second week of March 2027
Team size 7
Repository UndergraduateArtificialIntelligenceClub/Failure-Forensics
Backlog Jira project KAN · Confluence space FF

About the project

Failure Forensics is a project of the Undergraduate Artificial Intelligence Club at the University of Alberta. There is no external client: the club team defines the scope, builds the tool and presents it at the club demo in early March 2027. The goal is a working tool members can use to debug multi-step LLM applications (RAG pipelines, agents, document pipelines) and to learn how AI systems fail.

Open questions

The blueprint uses the abbreviation UAIS in its headers and mentions a decision model called Jev. Both are tracked in the Glossary.

The problem

Modern AI applications run many steps: rewrite the question, retrieve documents, rerank them, build a context, call a model, call tools. When the final answer is wrong, every step usually reports success. Logs show that something failed, not where it started or why.

Teams end up reading raw traces by hand, guessing at the cause, fixing one prompt, and having no way to check that the same failure does not come back.

What we're building

A closed loop for AI failures:

flowchart LR
  T[Trace] --> D[Detect] --> G[Diagnose] --> E[Explain] --> V[Verify] --> L[Learn] --> R[Regress] --> I[Improve]
  I -.-> T
  1. A Python SDK that records every step of an AI pipeline as spans, including which step consumed which step's output.
  2. A backend that stores traces and runs detectors (rules, statistics, a DecisionGate classifier and LLM judges).
  3. A forensics engine that walks backward from the bad answer to the earliest step that degraded the data, and labels downstream steps as propagated.
  4. An evidence and grounding gate so every claim in a diagnosis points at stored trace data.
  5. A human review step that turns confirmed failures into regression test cases.
  6. A Next.js developer console that shows the trace graph coloured by root cause, propagation and health.
  7. Benchmark apps (document, RAG, agent) with controlled fault injection to measure root-cause accuracy.

Who it's for

User What they need
AI / ML engineer Find which step of an LLM, RAG or agent pipeline broke.
Backend engineer Understand failures that cross APIs, tools, prompts and transforms.
Platform / DevOps engineer Watch latency, cost, errors and worker health.
AI product team Track quality, failure categories and release regressions.
Research / evaluation engineer Build datasets, inject faults and measure diagnostic accuracy.

Stakeholders

Stakeholder Interest
Undergraduate AI Club A working, demo-able tool built by and for club members.
Development team (7) Shipping the MVP by 29 Jan and code-complete by 26 Feb.
Club members and future contributors Clear docs, ADRs, tests and a reproducible local setup to keep building on it.

Team

Role Owner Focus
Tech Lead & Architect T1 Scope, contracts, ADRs, integration, demo
SDK & Ingestion T2 Python SDK, span links, ingestion, compare view
Backend & Infrastructure T3 Database, APIs, jobs, security, CI/CD, staging
AI Detection T4 Taxonomy, validators, DecisionGate, LLM judges
AI Forensics T5 Root-cause engine, propagation, evidence, grounding
Frontend / UX T6 Developer console
Evaluation & Benchmarks T7 Benchmark apps, fault injection, eval runner, report

Names are added on the Roles page once the team confirms them.

Live app

The staging deployment goes live in Sprint 7 (from 1 Feb 2027). Until then, see Running Locally.

Documentation