Inspiration

Cloud-based AI research tools routinely expose sensitive IP to third-party servers, hallucinate citations, and fracture study workflows across disconnected tools like Notion, Anki, and standalone PDF readers. Furthermore, passive reading leads to rapid retention drop-off within 48 hours. We built Knowledge OS to give researchers, students, and engineers a sovereign, 100% local AI cockpit that turns static source papers into active recall, mathematical proofs, and rigorous oral defense mastery.

What it does

Knowledge OS ingests PDFs, Office documents, and YouTube lectures into a local, multi-pane research workspace:

Autonomous 3-Agent Research Team: A Scout gathers verified empirical web sources, a Fact-Checker evaluates consensus and confidence scores, and an Academic Lead Author drafts structured literature reviews with verified citations.

AI Mock Viva & Defense Studio: Simulates doctoral oral exams using adjustable personas (Strict Board Member, Academic Mentor, Socratic Inquisitor) with automated 4-pillar rubric scoring.

STEM Formula Engine: Automatically extracts LaTeX equations, maps physical variable glossaries, deconstructs 3-step first-principle derivations, and runs Python Matplotlib/NumPy simulations.

Active Learning Hub: Features dual-host NotebookLM-style audio podcasts, 3-panel Cornell notes with slash-command co-writing, 1-click PowerPoint generation, and SM-2 spaced repetition with Anki export.

Epistemic Concept Diff: Compares two conflicting research papers side-by-side to highlight agreements, contradictions, and novel contributions.

How we built it

Local AI Core: Powered entirely on-device with Ollama running qwen2.5:3b and nomic-embed-text, ensuring zero cloud data leakage.

Vector & Storage Engine: Thread-safe FAISS CPU indexing with normalized inner-product retrieval, backed by FastAPI (Python 3.11) and SQLite/PostgreSQL with multi-tenant isolation.

Frontend Cockpit: Next.js 15 App Router, TypeScript, and Tailwind CSS, featuring an interactive force-directed knowledge graph and Server-Sent Events (SSE) for streaming RAG responses.

Document Processing: PyMuPDF, pdfplumber, and python-docx with sliding-window chunking (800 tokens, 150-token overlap) and verified page citation tracking.

Challenges we ran into

Local Hardware Optimization: Orchestrating multi-agent web searches, fact-checking consensus, and RAG retrieval smoothly on consumer hardware without GPU-bound latency.

Preserving Mathematical Context: Designing deterministic chunking boundaries that maintain the semantic integrity of LaTeX equations and slide presentations.

Defeating Sycophancy in Viva Mode: Tuning the defense simulator prompts so the examiner maintains realistic academic skepticism and methodological critique rather than defaulting to agreeable praise.

Accomplishments that we're proud of

100% Pass Rate: Architected a robust test suite with 21/21 unit, integration, and end-to-end tests passing.

True Sovereignty: Zero external API dependencies required for core retrieval, synthesis, flashcard generation, or note taking.

End-to-End Pipeline: Delivering an integrated workflow from raw document ingestion all the way to oral viva defense and native .pptx presentation generation.

What we learned

Compact local language models (like Qwen 2.5 3B) deliver exceptional domain depth when backed by tight semantic retrieval, strict prompt scaffolding, and specialized agent roles.

Active Socratic pressure testing and oral defense simulations provide significantly higher conceptual retention than passive summarization tools.