• 

• 

• 

• 

🤖 Casjoe Local AI — 100% Offline Business Intelligence & Autonomous Agent Engine

Throughput on standard 4-vCPU hardware achieves generation speeds of \( \Phi_{\text{gen}} \ge 24.8 \text{ tokens/sec} \) with a Time-To-First-Token (TTFT) of \( \text{TTFT} \le 0.42 \text{ s} \).

⚠️ Challenges We Ran Into

1. Client-Side RAG Without Server Backends: Standard RAG pipelines use heavy Python backends. We engineered a custom in-browser extr## 💡 Inspiration

Across Africa and developing regions worldwide, over 80% of employment is driven by micro, small, and medium enterprises (MSMEs). Yet, these entrepreneurs, clinic managers, and freelancers face severe barriers to adopting modern AI:

1. Unreliable Internet & Blackouts: Frequent power grid load shedding and costly cellular data make cloud-based AI tools (e.g. ChatGPT Plus, Claude, cloud APIs) impractical or unusable.
2. Data Sovereignty & Privacy Concerns: Small businesses handling client invoices, medical triage records, or financial projections cannot afford to transmit sensitive records to foreign cloud servers.
3. The Prompt Engineering Gap: Non-technical operators often struggle to structure effective prompts needed for actionable enterprise outputs.

We built Casjoe Local AI to solve these challenges. Our vision is to democratize enterprise-grade AI by bringing a 100% offline, privacy-first, zero-subscription AI business engine directly onto budget consumer hardware (8 GB RAM laptops)—without requiring a single byte of internet connectivity.

⚡ What It Does

Casjoe Local AI is a complete, local-first enterprise management workspace powered by quantized local Large Language Models (Llama 3.2 3B GGUF via llama.cpp / Ollama):

• 🔌 100% Offline AI Chat: Chat, brainstorm, and generate business strategies locally with zero network latency.
• 🤖 Autonomous Agentic Actions: Type natural commands like "Create an invoice for Sarah for $400 for website design services" or "Add customer Sahara Logistics to CRM", and the local AI automatically parses the parameters and updates the local database.
• 📄 Client-Side Document RAG (PDF & TXT): Drag and drop proposals or contracts into the workspace. Text extraction and retrieval run 100% in-browser via web workers without uploading files anywhere.
• 📚 120+ Sector Prompt Library: Pre-loaded expert prompts across 12 sectors (Business, Healthcare, Legal, Finance, Marketing, IT, Real Estate, HR, E-Commerce, Education, Sales, Customer Support) tailored for African enterprises.
• 💼 Full Business Management Suite: Integrated CRM, Invoicing & Finance Tracker, Inventory Manager, Point of Sale (POS) terminal, and Document Vault using local IndexedDB storage.

🛠️ How We Built It

We engineered Casjoe Local AI using a modern, privacy-focused stack built for low-resource hardware:

• Frontend UI & Styling: React 19, Vite, and TailwindCSS featuring a dark gold enterprise UI theme.
• Inference Engine & Quantization: llama.cpp GGUF runtime running Llama-3.2-3B-Instruct quantized at Q4_K_M (4-bit medium with 6-bit critical tensor quantization).
• Desktop Packaging: Electron and electron-builder producing native Windows .exe installers and Linux .tar.gz / .zip binaries.
• Offline Database: dexie (IndexedDB) for local relational persistence of customer, invoice, and inventory records.
• Client-Side RAG & PDF Export: pdfjs-dist for offline vector/text parsing and jspdf / html2canvas for offline PDF generation.
• Data Visualization: Chart.js (react-chartjs-2) for dynamic revenue and financial analytics.

📐 Mathematical & Memory Model

To guarantee smooth performance within the strict 8 GB RAM laptop profile, we benchmarked token latency \( T_{\text{latency}} \) and RAM memory allocation \( M_{\text{total}} \):

Mtotal=MOS+Mruntime+Mweights+Mcontext

Given:

• \( M_{\text{weights}} \approx 2.02 \text{ GB} \) (for Llama-3.2-3B Q4_K_M)
• \( M_{\text{context}} \approx 0.80 \text{ GB} \) (for 4K KV-cache context window)
• \( M_{\text{runtime}} \approx 0.40 \text{ GB} \) (Electron + React 19 shell)

The total operational RAM bound stays strictly within:

$$ M_{\text{total}} \le 4.22 \text{ GB} < 8.00 \text{ GB} $$

action pipeline using pdfjs-dist inside Vite web workers to prevent UI freezes on low-spec laptops.

1. Deterministic Agentic Tool Usage on 3B Models: Compact 3B models can struggle with complex function calls. We built structured prompt guards and fallback parsers that achieve high tool-calling precision without adding latency.
2. Cross-Platform Packaging: Configuring Electron builders to bundle local SIMD-optimized binaries across both Windows and Linux without missing dynamic link libraries.

🏆 Accomplishments That We're Proud Of

• 🚀 Zero Internet Required: From model inference to PDF invoice rendering and CRM updates, 100% of features run offline.
• ⚡ High Throughput on Budget Laptops: Sub-second response initiation on 8 GB RAM machines with low CPU thermal impact.
• 🌍 African Enterprise Tailoring: Multi-currency support (NGN ₦, GHS GH₵, KES KSh, ZAR R, RWF FRw, EGP E£) and 120+ pre-built sector workflows.

📚 What We Learned

• Quantized Edge LLMs are Production-Ready: Llama-3.2-3B-Instruct at Q4_K_M delivers near-lossless instruction accuracy for business tasks while preserving over 50% system RAM.
• Local-First Architecture Powers Resilience: Combining browser storage (IndexedDB) with local GGUF models creates an extraordinarily fast and resilient UX unaffected by ISP outages.

🚀 What's Next for Casjoe Local AI

• 📱 Android & Mobile Deployment: Porting GGUF execution to mobile devices (via llama.cpp Android bindings) for market vendors.
• 🎙️ Offline Voice Assistants (Whisper): Integrating local offline speech-to-text for hands-free operation in regional African languages.
• 🔗 Peer-to-Peer Local Mesh Sync: LAN/Wi-Fi database synchronization between local shop devices without touching external cloud servers.

Built With

• africa-tech
• autonomous-agents
• chartjs
• desktop-app
• dexie-js
• electron
• gguf
• indexeddb
• javascript
• jspdf
• laptop-llm
• llama.cpp
• llama3.2
• node.js
• offline-ai
• ollama
• pdfjs
• privacy-first
• rag
• react
• sme-tech
• tailwindcss
• vite