Heirloom | Devpost

Inspiration

At ADP, important knowledge can live in the heads of the people who use it every day. When someone changes teams or leaves, that context can disappear with them. At the same time, teams working on related problems may not realize they are duplicating each other’s work.

We wanted to build something that treats organizational knowledge as something that should be preserved, maintained, and shared, rather than something people have to repeatedly search for.

That led us to Heirloom: a system where every team has its own AI “Keeper” that understands the team’s knowledge, learns from its experts, and communicates with other team agents when a question crosses organizational boundaries.

What it does

Heirloom is an AI-powered organizational knowledge system designed to make sure important knowledge outlives the people who hold it.

Each team gets its own AI Keeper that:

Answers questions using the team’s documented knowledge and provides source-backed answers.
Communicates with other team Keepers when a question requires cross-team context.
Routes unanswered questions to the right human expert.
Turns expert responses into structured knowledge that can be reviewed and approved.
Identifies knowledge that is at risk of being lost through a bus-factor map.
Runs proactive “agent standups” to identify duplicated work, opportunities for reuse, and conflicts between teams.
Performs impact checks when a team proposes a change that could affect other teams.
Creates ramp-up paths for new team members.

The key idea is that Heirloom doesn't just search what an organization has already documented. It helps capture what people know but haven't documented yet, then makes that knowledge available to the people who need it.

How we built it

Heirloom uses a full-stack architecture with a Next.js frontend and FastAPI backend.

The frontend provides the interactive workspace for asking questions, reviewing knowledge, viewing the agent network, managing expert questions, and exploring organizational insights.

The FastAPI backend acts as the orchestrator. It routes questions to the appropriate team agent, manages agent-to-agent calls, handles expert routing, verifies citations, and streams live status updates to the frontend.

Each team has an independent AI Keeper built from the same agent codebase but scoped to that team's knowledge. The agents communicate through MCP tools, allowing them to ask one another questions while respecting team-specific sharing permissions.

We use Supabase Postgres with pgvector for structured data and semantic retrieval, with Supabase Storage for uploaded documents and voice notes. Documents are parsed, chunked, and embedded for retrieval. AI is used for answering questions, structuring expert responses, generating insights, and assessing cross-team impact.

For the demo, we use mock Payroll, Benefits, and Time & Attendance teams and intentionally planted scenarios including knowledge gaps, duplicated work, conflicting documentation, and restricted information.

Challenges we ran into

The biggest challenge was making an AI system trustworthy enough for organizational knowledge.

A normal chatbot can produce a convincing answer even when it does not actually know the answer. We wanted Heirloom to behave differently.

We designed the system around several safeguards:

Answers are labeled based on their certainty.
Sources are passed through to the final response.
Citations can be verified against the original source.
Conflicting expert answers are surfaced instead of silently resolved.
Unanswered questions are routed to humans.

We also had to think carefully about how multiple team agents communicate. Rather than giving one AI access to everything, each Keeper has an explicit knowledge boundary and sharing policy. Agents can ask other teams for information, but they cannot access information that their team has not permitted them to share.

Finally, we had to balance the scope of the idea with a five-minute hackathon demo. We prioritized the core knowledge-capture loop, cross-team agent communication, and proactive insights so the demo would show a complete working story rather than a collection of disconnected features.

Accomplishments that we're proud of

We are especially proud of turning the idea of “organizational memory” into an interactive system rather than just another document search tool.

The most important part of Heirloom is the learning loop: when the AI cannot answer a question, it identifies the appropriate expert, captures their answer, structures it into a knowledge proposal, and allows a knowledge owner to approve it. The next person who asks the same question can then get an immediate, verified answer.

We are also proud of the agent network. Instead of one AI having unrestricted access to every team’s information, each team has an agent with its own knowledge and permissions. Agents can collaborate when needed while preserving ownership and visibility into what was shared.

Finally, our proactive insights demonstrate a different use of AI. Heirloom can compare what different teams are working on and surface potential overlaps, reuse opportunities, and conflicts before someone explicitly asks about them.

What we learned

We learned that building a useful AI product is less about simply adding an LLM and more about designing the system around the model.

For organizational knowledge, trust, ownership, permissions, provenance, and human review are just as important as the model's ability to generate an answer.

We also learned that separating knowledge by team can actually make an AI system more useful. Instead of creating one massive knowledge base with unclear ownership, each Keeper has a clear domain and responsibility. That makes it possible to ask not only “What does the company know?” but also “Who owns this knowledge?” and “Who should I talk to next?”

Most importantly, we learned that the most valuable AI systems may not just answer questions. They can identify missing knowledge, connect people who should be talking, and prevent problems before they happen.

What's next for Heirloom

The next step would be connecting Heirloom to the systems where organizational knowledge actually lives, such as Microsoft Teams, Outlook, Slack, project management tools, meeting transcripts, and internal documentation.

We would also expand the knowledge lifecycle with stronger versioning, richer permission controls, better correction propagation, and analytics showing how much time teams save by reusing existing knowledge.

Ultimately, we want Heirloom to become an organizational memory layer: a system that continuously captures what teams learn, connects knowledge across organizational boundaries, and makes sure important context survives changes in people, projects, and teams.