Inspiration

  • The Knowledge Transfer Gap: At my own company, we experienced significant friction during Knowledge Transfer (KT) despite having ample resources; the process remained slow and fragmented.
  • The Onboarding Bottleneck: Research shows that in large Multi-National Corporations (MNCs), the onboarding and KT process typically consumes anywhere from 2 weeks to 12 weeks.
  • The Knowledge Graveyard: Enterprises own decades of documentation, PDFs, and Slack threads that no one reads, leaving new hires to drift while critical information remains buried.

What it does

  • Proactive Briefing: Upon selecting a role, the agent automatically reads all company documents to generate a personalized 30-day roadmap before the user even asks a question.
  • Role-Aware Synthesis: The system reframes answers based on the user's department; an engineer receives technical commands while a marketer is directed toward high-level policies.
  • Tribal Knowledge Recovery: By indexing messy Zoom transcripts alongside formal docs, it captures the "why" behind technical decisions that never made it into the official wiki.

How we built it

  • 12-Hour Sprint: Developed using a Python-based RAG (Retrieval-Augmented Generation) system designed for high-speed delivery during a hackathon.
  • Streamlined Tech Stack: Built with a Streamlit UI for the frontend and FAISS for a high-performance, in-memory knowledge base.
  • Dual-Agent Architecture: Utilized two distinct OpenAI-powered agents one for proactive briefing and one for grounded, citation-heavy search.

Challenges we ran into

  • Formatting Rigidity: Early versions provided blunt paragraphs; we had to rewrite prompts to enforce structured markdown and verbatim italics for technical accuracy.
  • Role Accuracy: Engineering the system so the AI accurately changed its "voice" and filtered content appropriately for specific departments was a complex task.
  • Hallucination Guards: We implemented strict "I don't know" instructions to ensure the AI would not invent company policies when information was missing from the source docs.

Accomplishments that we're proud of

  • Zero-Setup MVP: Created a functional, role-aware tool that works in 60 seconds just by uploading existing documents.
  • Verbatim Citations: Developed a citation engine that copies exact commands, names, and URLs directly from the source to ensure 100% technical reliability.
  • Unified Source of Truth: Successfully combined data from messy handbooks, architecture docs, and meeting notes into a single, intelligent response.

What we learned

  • Focus Wins: Building a specific "Onboarding Copilot" was a much more compelling story and effective solution than a general "Enterprise Search" tool.
  • Prompting is Engineering: We learned that prompt engineering was the most critical part of the build, proving to be just as essential as the backend logic.

What's next for Day1 Brain

  • Live Connections: Moving from static file uploads to direct integrations with platforms like Slack, Jira, and GitHub.
  • Conflict Detection: Adding a "Conflict" flag to alert users when two different documents provide contradictory policy information.
  • Freshness Badges: Refining how the system surfaces document "freshness" so users can identify when an answer is based on outdated sources.

Built With

  • faiss
  • fastapi
  • next.js
  • openai
  • python
  • streamlit
Share this project:

Updates