-
-
Every incident leaves evidence, but scattered reports, logs and conflicting statements make finding the truth slow.
Inspiration
Every incident leaves evidence, but finding the truth takes time. Investigators often have to dig through scattered reports, emails, logs and photos. They compare conflicting statements and notice missing information only late in the process. Manual investigation is slow, inconsistent and hard to audit. We wanted to build an assistant that does the tedious cross-checking and leaves the judgment to the investigator.
What it does
TracePilot AI is an evidence-driven incident investigation assistant. An investigator uploads the case material: reports, logs, emails, statements and notes. TracePilot then:
- Extracts events from every source and reconstructs a timeline
- Cross-checks sources and flags conflicting statements (for example, a supervisor notified at 02:10 according to an email, but 02:25 according to a statement)
- Highlights missing evidence, such as a gap in CCTV coverage
- Drafts findings, each with source citations, a confidence level and a recommended follow-up
The investigator stays in control. Every finding can be approved, edited or rejected, and every action is written to an audit trail. The result exports as a clear, evidence-backed report.
How we built it
- Reasoning: NVIDIA Nemotron, served through Nebius Token Factory
- Backend: FastAPI (Python) handles ingestion, prompt orchestration and the review workflow
- Frontend: React and TypeScript, with an evidence panel, timeline, finding detail view and review controls
- Storage: SQLite holds evidence, findings, reviewer decisions and the audit log
The key design choice is that every finding links back to the exact source excerpt it came from. Nothing is presented as a conclusion without traceable evidence.
Challenges we ran into
- Getting the model to return consistent, structured output (events, conflicts and findings) that the UI could render reliably
- Keeping every claim grounded in a citation rather than a plausible-sounding summary
- Separating real conflicts between sources from harmless differences in wording
- Designing the review flow so the AI speeds the investigator up without taking over the decision
Accomplishments that we're proud of
- A complete workflow from raw evidence to a reviewed, exportable report
- Citations on every finding, so conclusions can be checked
- A human-in-the-loop review step with a full audit trail
- Running the whole system on a simple, portable stack
What we learned
- Trust in AI comes from traceability and human control, not just model quality
- Structured prompting and output validation matter as much as the model choice
- Good investigation tools highlight uncertainty (conflicts, gaps, confidence levels) instead of hiding it
What's next for TracePilot AI
- Support for more evidence types, such as images, video metadata and scanned documents
- Multi-investigator collaboration and role-based review
- Stronger conflict detection and confidence scoring
- Integrations with existing case-management and ticketing tools
- Evaluation on real-world incident datasets to measure accuracy and time saved
Built With
- ai-agents
- fastapi
- generative-ai
- incident-investigation
- llm
- nebius
- nebius-token-factory
- nvidia-nemotron
- python
- react
- sqlite
- typescript
Log in or sign up for Devpost to join the conversation.