Kollab

Inspiration

Cross-disciplinary teams lose the why in the handoff. A scientist’s finding gets retyped into a ticket, the ticket drops the constraint that made it necessary, and an engineer ships something subtly wrong. Chat and docs scatter context; AI chatbots regenerate it from scratch. We wanted one place where people and agents could Kollab: collab on the same canvas, with citations still attached.

The Mistral × Atlassian challenge asked how AI can help teams make sense of information, present ideas, align on decisions, and review work. Kollab is our answer: not another chatbot, a shared semantic board where research becomes assignable work agents can also read.

What it does

Kollab is one canvas for multi-disciplinary teams and AI.

Make sense: Drop a PDF, whiteboard photo, or CSV. Mistral proposes findings and constraints (with quotes/pages). Humans pin, summarize, or re-steer Sense.
Align: Spot contradictions across nodes, record decisions (decided / deferred / rejected).
Present: Stakeholder Present Mode plus a generated one-pager.
Review: Assign into Jira, pull PR lineage over MCP, run a constraint checklist against the work.
Recommend: From a finding or constraint, auto-create linked tasks grounded in the graph.

How we built it

Frontend: Next.js canvas (React Flow), pixel-brutal UI, Sense review, Present Mode, lineage focus, My Tasks.
Backend: FastAPI + SQLite as system of record (nodes, edges, candidates, activity).
AI: Mistral-only via the official SDK: OCR/chat for Sense, Agents/Conversations for the specialist roster (Coordinator, Sense, Data, Arbiter, Narrator, Reviewer), image generation for one-pagers.
Integrations: Jira outbound on assign; GitHub/Jira webhooks; MCP tools so a coding agent can get_task_context and report a PR back onto the board.
Lineage: Graph walk, not a prompt, so agents and humans see the same ancestry.

Challenges we ran into

Keeping eligibility strict: every reasoning step on Mistral, with Cursor only as an MCP client.
Trust: auto-extract is noisy. We built candidate review, re-sense with notes, append-by-revision, and “summary only” so the canvas doesn’t flood.
Merging design + product under time pressure (pixel retheme, teammate UX, Sense packets) without breaking API contracts.
Making Present / Align / Review feel like one story instead of disconnected demos.
How to make it unique? Canvas it's not something new, but the value over agent communication is never there.

Accomplishments that we're proud of

A live end-to-end path: PDF → Sense → promote → align → present → Jira → MCP PR write-back.
Human-in-the-loop Sense that still feels fast.
Real lineage for agents and people on the same graph.
A clear product name and pitch: Kollab = collab on one canvas.

What we learned

The hard part isn’t “call an LLM”; it’s where truth lives. The canvas had to stay the system of record.
Multimodal ingest only wins if citations survive into tasks and tickets.
Specialist agents map cleanly to HMW (HOW MIGHT WE?) verbs when graph tools, not chat history, are the write path.
Design readability (weight, contrast, when to use red) matters as much as features in a live demo.