Elisha622/omnicore-ai

OmniCore AI

An enterprise multimodal AI workspace — chat, document intelligence, computer vision, speech, AI agents, churn prediction, enterprise search, automation, integrations and trade-regulation research. Open it and it works: no server to run, no Docker, no model download, no API key. Sign in once with a free Hugging Face account and the AI runs on Hugging Face's servers, or switch to a fully private on-device mode. Installable as an app.

Live app

• Hugging Face Space: https://huggingface.co/spaces/Elisha622/omnicore-ai
• GitHub Pages: https://paulelisha500-ops.github.io/omnicore-ai/

What's inside

Module	What it does
Executive Dashboard	KPIs, a 14-day activity trend and module usage computed from your real activity, plus an on-demand AI insight.
AI Chat Assistant	Saved conversations with streaming answers, grounded in your knowledge base (with citations), optional Wikipedia lookup, image questions.
Document Intelligence	PDF, Word, text, CSV, HTML or a photo of a page → summary, key points, document type and Q&A. Long documents are summarized part by part; scanned pages are read by the vision model.
Computer Vision	Describe a scene and its objects, read text in an image (multilingual), or ask your own question. Works with the phone camera.
Speech AI	Private on-device transcription with Whisper (record or upload), meeting summaries with decisions and action items, text-to-speech with voice/speed/pitch.
AI Agent Platform	Eight agents. Research pulls live sources and cites them, the Data Analyst profiles a CSV you attach, the Reporting Agent reads your real workspace numbers.
Predictive Analytics	A real scikit-learn Random Forest churn model (IBM Telco, 7,043 customers) running in the browser: live what-if predictions, risk-lowering levers, and batch scoring of whole CSV files with downloadable results.
Recommendation Engine	Next steps, related documents, one-click automations and at-risk customers, all derived from your activity.
Enterprise Search	Instant ranked search across chats, documents, meetings, analyses, agent runs and automations, with an AI answer from the results.
Trade & Regulatory	Live retrieval from official sources (US Federal Register, GOV.UK, EU Open Data Portal), per-authority site search, a verified directory, and strictly cited AI summaries.
Automation Platform	"When this happens, do that" rules on real app events and schedules — notify, post to Slack/Discord, call a webhook, open an email, or save to the knowledge base.
Integration Hub	Slack, Discord, custom webhooks, email and desktop notifications, each with a one-click test.
Security & Governance	Accounts with PBKDF2-hashed passwords, sign-in lockout, enforced roles, user management, and a filterable, exportable audit log.

Everything is bilingual (English and Arabic, with full right-to-left layout) and has light and dark themes.

How it works

• Online AI (default). "Sign in with Hugging Face" (OAuth on the Space, inference-api scope) gives the app a token for Inference Providers: GPT-OSS 20B for text and Qwen3-VL 30B for images, answering in about a second. Usage is billed to each visitor's own Hugging Face account, whose free monthly credit covers everyday use. Nothing downloads.
• On-device AI (optional, private). Settings → AI engine → On this device: Transformers.js runs Qwen3 0.6B (text), Qwen3.5 (images) and Whisper (speech) in a Web Worker with WebGPU, after a one-time download. Nothing you type leaves the device.
• Your data stays local. Accounts, documents, chats, automations and history live in the browser's IndexedDB. Export/import/erase are in Settings.
• Installable app. It's a Progressive Web App: use "Install app" in the top bar (or Share → Add to Home Screen on iPhone).
• Real churn model. ml/export_churn_model.py exports the trained scikit-learn pipeline to frontend/public/models/churn-forest.json; the browser evaluator matches scikit-learn's predict_proba to within 1e-6.

Requirements

Any current browser. Online AI needs a free Hugging Face account (one-click sign-in) and works on the Hugging Face–hosted app. On-device AI is fastest with a WebGPU-capable GPU.

Run it locally

cd frontend
npm install
npm run dev

Then open http://localhost:5173. Build a production bundle with npm run build (output in frontend/dist).

Deploy

Build once with cd frontend && npm run build, then from the project root:

• GitHub Pages: bash scripts/deploy_gh_pages.sh publishes frontend/dist to the gh-pages branch (Pages source: gh-pages, /).
• Hugging Face Space: python scripts/deploy_hf_space.py (needs hf auth login).

Retrain the churn model

python ml/train_churn_model.py      # trains and saves ml/churn_model.joblib
python ml/export_churn_model.py     # exports to the browser and verifies against sklearn

Project layout

frontend/src/
  App.jsx              app shell: auth gate, navigation, AI status, sheets
  AuthScreen.jsx       create account / sign in
  LandingPage.jsx      public landing page
  modules/             one file per screen
  lib/                 ai (worker + client), db, auth, data, search, web,
                       docs, churn, automation, integrations, trade, audio
ml/                    churn model training + export
scripts/               Hugging Face Space deploy

See PROJECT_NOTES.md for design decisions and honest limits.