DriftGuard

Inspiration

Modern software teams don't struggle because they lack information—they struggle because they have too much of it.

A single feature can span GitHub pull requests, architecture documents, product specifications, meeting notes, and design discussions. As these sources evolve independently, they inevitably drift apart. Product managers, designers, and engineers often end up working from different versions of the truth, leading to misunderstandings, duplicated work, and costly implementation mistakes.

We wanted to solve this problem by creating a shared, evidence-backed understanding of a project that adapts to each person's role instead of forcing everyone to interpret technical information the same way.

What it does

DriftGuard continuously analyzes project artifacts across documentation, code, and meeting decisions to detect semantic drift before it becomes a communication problem.

Instead of simply summarizing information, DriftGuard:

Ingests project documentation, GitHub repositories, pull requests, and meeting transcripts.
Detects inconsistencies between implementation, documentation, and recorded decisions.
Produces a single alignment view of the project backed by verifiable evidence.
Translates technical changes differently for product managers, designers, and developers.
Generates evidence-backed documentation updates and change logs for review before anything is modified.

Every insight links directly back to its original source, giving teams confidence in why a recommendation was made.

How we built it

We built DriftGuard using Mistral AI as the core reasoning engine, combining multiple capabilities throughout the pipeline.

Our system consists of several stages:

Project ingestion
- GitHub repositories and pull requests
- Google Docs documentation
- Meeting transcripts
Evidence extraction
- Documents are split into structured evidence spans.
- Code changes are converted into implementation claims.
- Meeting transcripts are converted into structured decisions.
Semantic reconciliation
- Mistral compares claims across every source.
- Contradictions, stale documentation, undocumented implementation, and unresolved decisions are identified.
Role-aware translation
- The same underlying project state is presented differently for engineers, designers, and product managers.
- Every explanation remains grounded in the same evidence.
Human-in-the-loop review
- DriftGuard proposes documentation updates rather than applying them automatically.
- Teams review, approve, and audit every change.

Challenges we ran into

The biggest challenge was determining what should actually be considered "truth."

Code, documentation, pull requests, and meeting discussions all represent different kinds of knowledge. A merged pull request doesn't necessarily represent approved product intent, while meeting discussions often contain proposals that were never accepted.

Rather than assuming one source is always correct, we built an evidence-first pipeline where every conclusion is supported by citations and every proposed change requires human approval.

Another challenge was balancing automation with trust. We wanted DriftGuard to save teams time without becoming another opaque AI assistant, so every recommendation is fully traceable back to the original documents or code.

Accomplishments that we're proud of

Building an end-to-end semantic drift detection pipeline powered entirely by Mistral.
Creating role-aware explanations that present the same project differently depending on the audience.
Designing an evidence-backed review system instead of a traditional black-box AI summarizer.
Integrating documentation, code, and meeting decisions into a single alignment workflow.
Delivering a working prototype that demonstrates how multidisciplinary teams can stay aligned as projects evolve.

What we learned

This project reinforced that the hardest part of collaboration isn't generating information—it's maintaining shared context.

We also learned the importance of structured outputs, evidence grounding, and human review when building AI systems that teams can trust. Rather than replacing decision-making, AI is most valuable when it helps people understand why something has changed and who it affects.

What's next for DriftGuard

Our roadmap includes:

Live integrations with Jira, Confluence, Slack, and Google Workspace.
Continuous monitoring of repositories and documentation through webhooks.
Automatic dependency graphs showing how one decision affects downstream work.
Richer collaboration workflows with approval routing and notifications.
Enterprise-scale deployment across multiple repositories and projects.

Ultimately, we want DriftGuard to become the shared source of truth that keeps multidisciplinary teams aligned as software projects grow in complexity.