• 

  Home Page - Dark Mode

• 

  Multi-Model Support

• 

  Research Assistant

• 

  Translator

• 

  Power Mode (Automation)

• 

  Memory Manager

SarvajñaGPT — Private, Agentic AI for Everyday Learning

SarvajñaGPT is an offline‑first, agentic AI that runs on your PC. It orchestrates specialized assistants (Research Assistant, Beginner‑Friendly Teacher, Homework Helper, Daily Life Companion) with a Memory Manager and local multi‑model control via Ollama—so you get focused help while your data stays on your machine.

Why this matters for education

• Personalized learning
  • Students get step‑by‑step explanations, breakdowns, and examples tailored to their level, without exposing their data to the cloud.
• Retrieval‑augmented answers from your own notes
  • Tag and organize your class notes, PDFs, and images. The app fetches just‑in‑time context from your tagged knowledge so responses are grounded.
• Better study workflows on Windows
  • Side‑by‑side layouts for Word/VS Code/Browser, selection‑aware enhancements (e.g., “explain this paragraph more clearly”), and a clean chat UI that renders math and code nicely.
• Inclusive learning
  • Built‑in translation and TTS help language learners and younger students, and create multiple representations of the same idea.

Privacy and security: local first by design

• Fully local by default
  • Models run via Ollama on your machine; no text is sent to a cloud LLM unless you explicitly choose to.
• Data never leaves your device
  • Your notes, chats, tags, and documents are stored locally. Clipboard‑first flows for Word changes keep control with you—no intrusive automation.
• Transparent operations
  • Clear logs and deterministic behaviors; you can inspect prompts, responses, and file operations in your own environment.
• Minimal permissions
  • No background services that phone home, no analytics by default; opt‑in only.

What inspired this project

I wanted the power of agentic AI—planning, tools, memory—without the privacy trade‑offs of typical web UIs. Students and professionals deserve:

• Reliable local workflows (code, documents, research)
• Fast iteration without vendor lock‑in
• Trust that sensitive notes and drafts stay private

How I built it

• Frontend: Vue 3 + Vite
  • Responsive chat + agent views, tag chips, attachment previews, mobile‑friendly input.
  • Power Mode panel for selection monitoring and clipboard‑first enhancements.
• Backend: FastAPI
  • Local inference bridge to Ollama (HTTP); model selection persisted to disk.
  • Endpoints for translation, document processing, and selection‑aware enhancement.
• Automation UX (Windows‑first)
  • A pragmatic approach to desktop assistance: reliable split layouts, clipboard‑first Word enhancements (user performs final paste to preserve formatting), optional CUA integration.
• Memory Manager + retrieval
  • Tag‑aware context assembly from local SQLite/embeddings. Folder‑level tags take precedence, promoting a tidy mental model.

Educational impact (examples)

• Homework Helper breaks down multi‑step problems with hints before solutions.
• Beginner‑Friendly Teacher rewrites dense passages into simpler language and adds scaffolding.
• Translator enforces native scripts and translation‑only outputs for clarity.
• Research Assistant composes answers that cite your notes and files, keeping study sessions grounded.

Challenges I faced

• Window focus and split heuristics on Windows
  • Maintaining stable tri‑split layouts (chat + editor + preview) and preventing focus flicker.
• Translation hallucinations
  • Models sometimes return explanations; we hardened prompts and added careful cleanup to ensure translation‑only outputs.
• Document editing without breaking formatting
  • Full‑document paste can trash styles; we shifted to snippet‑only clipboard replacements.

How it benefits teachers, students, and institutions

• Teachers
  • Generate differentiated materials, glossaries, and examples quickly—entirely offline on school hardware.
• Students
  • Keep notes private, study with targeted retrieval, and translate concepts without surveillance.
• Institutions
  • Deploy a local, privacy‑respecting assistant without data egress risks or ongoing per‑seat cloud costs.

Built With

• comtypes
• coqui-tts
• cua
• eslint
• fastapi
• gtts
• javascript
• lingua-language-detector
• node.js
• npm
• ollama
• ollama-(local-llm
• openpyxl
• pdf2image
• pdfminer.six
• pdfplumber
• pillow
• pydantic
• pymupdf
• pypdf2
• pypdfium2
• python-3.11
• python-docx
• python-pptx
• pytorch
• pywin32
• requests
• rest-api)
• sentence-transformers
• soundfile
• sqlite
• tailwind-css
• tesseract-ocr-(pytesseract)
• uvicorn
• vite-(esbuild/rollup)
• vue-3