Contextor

Inspiration

Modern teams rarely struggle because they lack information; they struggle because the same information means different things to different people. An engineering change might imply a validation rerun for a scientist, a filing change for a lawyer, and a delayed launch for an operations lead. We built Contextor to bridge those professional contexts. Out goal was not simply to summarise information, but to help every team member understand what shared project facts mean for them without losing the original source of truth.

What it does

Contextor is an n-way translator between professional contexts.

Project information is converted into a neutral, structured Intermediate Representation, or IR. Contextor then re-projects that shared record through each team member's self-described expertise, responsibilities, and interests.

With Contextor, teams can:

Enter project updates or ask questions through one conversational input.
Upload documents and extract proposed project facts.
Review and edit AI interpretations before saving anything.
Route proposed changes through an admin approval workflow.
Read a personalised summary or switch to the underlying IR evidence.
Trace generated claims back to the facts supporting them.
Detect contradictory facts when new information is merged.
Resolve conflicts by editing, discarding, or explicitly accepting both facts. The result is one shared record with multiple useful views, not multiple disconnected versions of the truth.

How we built it

We built the frontend with React, Typescript, Vite, Tailwind CSS, and shadcn components. The backend uses Python and Flask, with SQLite for persistent storage and Mistral for AI interpretation, document extraction, conflict detection, question answering, and personalised re-projection.

We deliberately separated AI interpretation from persistence. The model can propose changes, but nothing becomes part of the project record until a user reviews it and an administrator merges it.

Conflict detection runs whenever new information is merged. The newly recorded facts are compared with the existing IR, and genuine contradictions are stored for human resolution.

We also added server-side checks around model output. Citations must reference real entries in the project, conflict pair must contain valid entry IDs, and malformed operations are rejected before reaching storage.

Challenges we ran into

The biggest challenge was distinguishing a revision from a disagreement. If someone says a sampling rate changed from 1 kHz to 2kHz, the original fact should be updated. But if two teams independently assert incompatible requirements, silently replacing one would hide a real disagreement. We therefore modelled conflicts as their own persistent state.

Grounding generated explanations was another challenge. Asking a model to include citations was not sufficient , the model could still invent an ID or make a claim that exceeded its sources. We added server-side citation validation and kept the neutral IR directly accessible through the interface.

Accomplishments that we're proud of

We are especially proud that Contextor treats AI as an interpreter rather than an unquestioned database writer. Every change passes through human review, and merging remains a controlled operation.

Other accomplishments include:

Personalised explanations generated from one shared source of truth.
Traceable soruce markers connecting prose to neutral IR entries.
A complete two-stage review workflow.
Conflict detection and human-controlled resolution.
Profile-driven behaviour without hard-coded professional roles.

What we learned

We learned that useful AI collaboration requires more than a strong prompt. Reliability comes from designing the surrounding system: validation, permission, provenance, review stages, and clear boundaries around what the model can change.

We also learned that personalisation is partly about omission. A useful view does not merely rewrite every fact using different terminology; it identifies what matters to a particular person and filters out irrelevant detail.

What's next for Contextor

We also want to add:

A project timeline showing how facts and decisions evolved.
Richer evidence tracking for external documents and sources.
Integration with tools such as Slack, Jira, Notion, and document repositories.
Notifications tailored to each person's responsiblilities.
More expressive IR relationships between decisions, dependencies, risks, owners, and deadlines. The long term vision is for Contextor to become a shared reasoning layer for multidisciplinary teams: one trusted project record that every person can understand from the context in which they work.