Shepherd OS - Eclipse API Intelligence Pilot

Accessible Intelligence (Not Artificial)

Traditional AI Approach

Treats AI like a magic oracle:

Dumps massive context hoping AI "figures it out"
No business context - learns from scratch each time
Intelligence is artificial - disconnected from your team

Accessible Intelligence (Shepherd OS)

A harness for YOUR team's knowledge:

Mike knows CCCDTPL = CC fee → now everyone knows
Ops knows tagged nonstock rules → now it's a tool
Intelligence is REAL - from your people, made accessible

—

Your Expertise

Real knowledge from real people

—

The Harness

Tools that encode what you know

—

Made Accessible

Anyone can use the team's wisdom

Continuous

Learn, improve, share back

"The intelligence is YOUR intelligence and together let's make it accessible. What Mike knows about CC payments, what your ops team knows about nonstock orders - we build a harness that makes that expertise accessible and available to save everyone time. The LLM is just the interface. The intelligence are the human experts - You."

Data Quality: Humans Make Mistakes

Check numbers: "12032025C" vs "12102025CC" vs "12152025" (no suffix) vs "12202025c" (lowercase)

The intelligence layer catches what humans miss:

Detect: "CC payment has CCCDTPL but no suffix - data entry gap"
Flag: "Inconsistent format: C vs CC vs c - standardize?"
Learn: "CCCDTPL is the source of truth, suffix is optional indicator"
Improve: Feed patterns back to training and SOPs

AI Literacy: What's Actually in the Harness?

Demystifying what AI does (and doesn't do) in Shepherd OS

What the LLM Actually Does

Understands your question - "Show me CC payments" → knows you want the CC classifier tool
Picks the right tool - Sees 7 available tools, selects cc_payment_classifier
Formats the response - Takes structured data, presents it clearly
Handles conversation - Follow-up questions, clarifications

✗ What the LLM Does NOT Do

NOT the source of knowledge - It doesn't know CCCDTPL = CC fee
NOT searching raw data - No 765K token data dumps
NOT guessing business rules - Rules come from YOUR skills
NOT making decisions alone - Just orchestrates your encoded expertise

The Four Components of Shepherd OS

Skills

Markdown files with YOUR rules

Human-written, version-controlled

—

MCP Server

Tools that call Eclipse API

Efficient, targeted queries

LLM

The orchestrator/interface

Picks tools, formats output

Web UI

Where users interact

Simple, chat-based access

The Key Insight

AI is the messenger, not the message.
The LLM delivers your team's expertise to users who need it. Mike's knowledge about CC payments, your ops team's rules for nonstock orders - that's the real intelligence. The AI just makes it accessible.

Bad Analogy

"AI is a smart employee who knows everything"

This leads to dumping data and hoping for magic

Better Analogy

"AI is a skilled translator with a good toolbox"

You fill the toolbox with YOUR expertise, AI translates requests into tool use

▼ How It Works: The Architecture

Hover over any element to learn more

—

Human Experts

Mike: "CCCDTPL = CC fee"
Ops: "Type T = tagged nonstock"

Encode

◆

Skills (Markdown)

cc-payment-classifier
tagged-nonstock-finder

Become Tools

●

MCP Server (Eclipse)

Semantic tools → Eclipse API

Available To

LLM (Orchestrator)

"I see 7 tools... picking the right one"

Responds To

◉

Users

"Show me CC payments for S004284138"

↩ Feedback → Experts → Improve Skills ↻

Pricing & Deployment

Skills: The Continuous Improvement Engine

Skills are Markdown files - human-readable, version-controlled, and continuously improving:

1. Capture

Mike solves a problem → Document the insight as a skill

2. Encode

SKILL.md with triggers, rules, examples → Becomes a tool

3. Scale

Single source of truth → Available to all users instantly

                    skills/cc-payment-classifier/SKILL.md → "CCCDTPL = CC fee" + trigger keywords + examples
                

Deployment Options (SOW)

Two paths to Accessible Intelligence - both include cloud safety net. Professional booklets available for offline review.

Local-First + Cloud Safety (Recommended)

$27K-32K

Development: $20,000 • Hardware: $7K-12K

On-prem is primary • Cloud validates complex queries

AI Box for AI Harness ($7K-12K)
Local model (14B-70B+) - handles 95%+ of queries
Cloud backup for complex/edge cases
Dual verification = best accuracy
Data stays on-prem by default

Monthly costs breakdown:
• Electricity: $5-15/mo (AI Box uses ~9W idle, ~120W avg under load)
• Cloud validation API: $10-50/mo (only for complex queries)

Choose this if: Privacy matters, high query volume, want lowest per-query cost

Timeline: 10 weeks Data: Stays local by default

Cloud-First + Local Cache

$13,500

Development: $13,500 • Hardware: None required

Cloud is primary • Local caching for speed

No special hardware - use existing machines
Cloud APIs (production-grade LLMs)
Fastest deployment (8 weeks)
Local cache for common queries

Monthly costs:
• Cloud API usage: $0.002-$0.007 per query
• Est. 5,000 queries/mo = $10-35/mo
• Est. 20,000 queries/mo = $40-140/mo

Choose this if: Need it fast, lower upfront budget, flexibility to scale

Timeline: 8 weeks Data: Processed via cloud

All options include: Intelligent harness with semantic tools • Skills library (Markdown-based, version-controlled) • Continuous improvement workflow • Modern web UI • Model-agnostic architecture

Skills grow with your team: Your team adds skills over time - every solved problem can become a skill. No vendor dependency for updates. Source code owned by Shepherd Electric.

ShepherdOS

Good evening

CC Payment Classifier

Tagged Nonstock

Backlog Calculator

Utility Rebates

Order Lookup

Job Management

Innovo Files