🧠

ZEUS INTELLIGENCE

The AI Operating System: Graph Brain + RAG Brain + Math Engine + Causal Validator + Context Harness
2,808 RAG Chunks 54 Graph Nodes 10 Proven Formulas 5-Layer Trust Chain 97% Token Savings
📖 What is RAG?
R
Retrieval
Go FIND the relevant information from 2,772 chunks
A
Augmented
ADD that real BSP data to the AI's prompt
G
Generation
THEN generate a cited answer using real data
Without RAG, AI answers from general knowledge. It guesses.
With RAG, AI answers from BSP's actual playbooks, field notes, and data. It cites its sources.
💡 The Analogy
📚

Without RAG: A Library

You walk into a library with 98 books.
You need one specific answer.
You pull out 3-4 books, flip through hundreds of pages, read, compare, cross-reference.

Time: 30-60 minutes.
Accuracy: depends on which books you checked.
🧠

With RAG: A Genius

A genius sits at the desk who has memorized every page of all 98 books.
You ask your question.
The genius instantly recalls the 5 most relevant passages, synthesizes the answer, and tells you which book each fact came from.

Time: 10 seconds.
Accuracy: cites exact sources. Verifiable.
The genius gets smarter every day. Every new field note, estimate, and blog post adds to what Zeus knows.
⚙ How It Works (4 Steps)

YOU ASK

"What's the avg sewer ticket in Overland Park?"

🔍

ZEUS SEARCHES

Scans 2,772 chunks. Finds 5 most relevant in <1 second.

🧠

CLAUDE READS

AI reads 5 chunks. Writes answer using ONLY BSP data.

CITED ANSWER

"$1,782 avg [Source: Sacred HTML]. 126 jobs [Source: ST data]."

📊 What Zeus Knows
98
Documents
2,772
Knowledge Chunks
81MB
Institutional Knowledge
10s
Answer Time
🧠
ZEUS RAG
2,772 chunks
📚 98 Playbooks
2,632 chunks
📝 1,179 Field Notes
130 chunks
🔍 2,976 Keywords
10 clusters
💰 5,000 Estimates
$3.1M pipeline
📞 675 Calls
Attribution data
🔧 2,281 Pricebook
Items + pricing
📊 Proof: What Zeus Found in 98 Documents
8
Contradictions
16 different revenue numbers across docs
236
Action Items
Buried across docs, now all visible
75
Current Docs
Updated within 7 days
0
Stale Docs
Nothing older than 30 days
📋 View Full Knowledge Audit
👥 Who Owns What (Document Taxonomy)
👑 Stephanie33
P&L, CEO briefs, QB, membership, growth strategy
🔧 Kalen34
Pricing, revenue proof, brand, estimates, sewer expertise
🎨 Audrey55
Creative briefs, blog posts, brand, social, KSHB, content
📞 Ashton33
Operations, dispatch, estimates, field notes, invoicing
💻 Robert68
SEO, Google Ads, data weapons, Titan Killer, APIs, intelligence
👥 Team35
War rooms, standups, 100 Year, membership, community
⚡ How to Use Zeus
💬

Ask Zeus

Type any question about BSP. Get a cited answer in 10 seconds from 2,772 knowledge chunks.

Open Zeus Chat
📋

Knowledge Audit

See all 98 documents categorized. Find contradictions. 236 action items by owner.

Open Audit
📚

Document Library

Browse all 98 playbooks, briefs, reports, and strategies. Every doc linked and accessible.

Browse Library
🧬 The Intelligence Stack (NEW)
Five systems. One brain. No other plumbing company on earth has this.
🌐
Graph Brain
54 nodes, 60 edges
📚
RAG Brain
2,808 chunks
📐
Math Engine
10 proven formulas
🔗
Causal Validator
4 attribution chains
🎯
Context Harness
97% token savings
🌐 Graph Brain: The Nervous System
🌐

Every System Mapped

The Graph Brain maps every service, API, timer, database table, person, and KPI as a queryable dependency network.

Ask: "What breaks if the sync daemon fails?"
Answer: 9 systems cascade. Command Center, Dispatch Board, Tech Scoreboard, Custom Reports, Collections -- all go stale.
💥 CASCADE: titan-sync-daemon fails
command-center dispatch-board tech-scoreboard custom-reports collections titan.jobs titan.customers titan.technicians titan.estimates
Node Types: ⚙️ Services (4) • ⏰ Timers (8) • 🔗 APIs (9) • 💾 Tables (7) • 📱 Pages (4) • 👤 People (5)
🧠 ANALOGY: Like a city's electrical grid map. You don't guess which buildings go dark when a transformer fails -- you TRACE the circuit. The Graph Brain traces every data circuit in Nexus.
📐 Math Engine: Provable Numbers
📐
Every Number Has a Proof
Triple-format: LaTeX (proof) + Python (automation) + Excel (Stephanie)
💰 COST PER LEAD
$$\text{CPL} = \frac{\text{Ad Spend}}{\text{Conversions}}$$
📜 LaTeX
🐍 Python
📊 Excel
BSP Current: $46/lead ✔️ (range: $5-$500)
📈 RETURN ON AD SPEND
$$\text{ROAS} = \frac{\text{Revenue}}{\text{Ad Spend}}$$
📜 LaTeX
🐍 Python
📊 Excel
BSP Current: 34.7x ✔️ (range: 0.5x-100x)
⚖️ OPERATING MARGIN
$$\text{Margin} = \frac{R - COGS - OpEx}{R} \times 100$$
QB Q1 2026: -5.5% ⚠️
⚖️ DIMENSIONAL ANALYSIS
$ ÷ leads = $/lead ✔️
$ ÷ $ = ratio ✔️
$ + leads = ???
🔒 5-LAYER TRUST CHAIN
1️⃣

Source

Is the raw data real?

2️⃣

Query

Is the SQL correct?

3️⃣

Formula

LaTeX + Python + Excel

4️⃣

Cross-Val

QB matches ST?

5️⃣

Certified

Full provenance stamp

🧠 ANALOGY: Like NASA's triple-redundant flight computers. Three independent systems must agree on every calculation. If any diverge, the mission aborts before the rocket explodes. Our numbers work the same way.
🔗 Causal Validator: Attribution Truth
🔗
"Did Google Ads Actually Drive That Revenue?"
Traces every attribution claim link-by-link through the data chain
📊 Google Ads → Revenue
Ad Click → Landing Page → Phone Call → Job Booked → Completed → Paid
33% Integrity
⚠️ Invoice Paid link BROKEN (payment_collected = 0)
🤖 Daniel AI → Revenue
After-hours Call → AI Captures → Slack DM → Callback → Revenue
0% Integrity
⚠️ No verifiable links (external data needed)
🧠 ANALOGY: Like a forensic evidence chain. In court, evidence is only admissible if every handler is documented. If one link breaks, the evidence is thrown out. Our attribution works the same way -- every link must be verified or the claim is flagged.
🎯 Context Harness: 97% Token Savings

Without Harness

50K
tokens per prompt
Dump ALL context every time. CLAUDE.md, 40+ memory files, meeting notes. AI sifts through noise, misses the signal, breaks downstream systems.
✔️

With Harness

1.4K
tokens per prompt (97% savings)
Graph queried for dependencies. RAG queried for warnings. Only 5-8 relevant nodes injected. Zero noise. Maximum precision.
🧠 ANALOGY: Without the harness: a surgeon reading the entire medical encyclopedia before every operation. With it: a surgical briefing card showing ONLY the patient's history, procedure steps, and known risks. Same surgeon. 10x more precise.
🚨 Anomaly Detector: Vital Signs Monitor
6 health checks. Every hour. Catches problems before humans notice.
✔️
Sync Freshness
Data less than 1hr old
✔️
API Health
All 5 endpoints responding
✔️
RAG Integrity
2,808 chunks intact
⚠️
Zero Invoice Rate
63% -- ST workflow broken
✔️
Daily Job Count
Within normal range
⚠️
Graph Integrity
15 orphan nodes
🧠 ANALOGY: Like a hospital's ICU vital signs monitor. Heart rate, blood pressure, oxygen -- all watched 24/7. When a number goes outside the expected range, the alarm sounds BEFORE the patient codes. Our systems work the same way.
💡 How Math Becomes Understanding
Numbers without context are noise. The Math Engine doesn't just CALCULATE -- it TRANSLATES between the language of data and the language of decisions.
🌍 → 📊 → 💬 → 🎯
THE TRANSLATION PIPELINE
🌍
RAW DATA
1,301 jobs
11 completed
$27,949 revenue
📐
MATH ENGINE
Decompose
Validate
Cross-check
💬
STORY
"$2.1M hidden
in unclosed jobs.
Techs not clicking Done."
🎯
ACTION
Fix completion
workflow. Train
techs on iPad.
🚰

Decomposition = Pipe Tracing

When a customer says "my water bill doubled," a plumber doesn't guess. They trace the pipes. Meter to main. Main to branches. Branch by branch until they find the leak.

The Math Engine does the same thing with revenue. "Revenue dropped 30%." Don't guess. Decompose:

Revenue = Leads × Booking Rate × Avg Ticket

Which pipe is leaking? Leads held at 130. Booking rate held at 80%. But avg ticket dropped from $2,912 to $1,800. Found it -- we did 6 drain cleanings ($150 avg) instead of sewer repairs ($5,000 avg). The "leak" is in the job mix, not the volume.
REVENUE PIPE TRACE
💰 Revenue: $27,949
└ 🔧 Completed Jobs: 11
   └ 🔍 Inspection: $19,500 (1 job)
   └ 🚨 Sewer: $7,200 (2 jobs)
   └ ⚡ Emergency: $982 (4 jobs)
   └ 🚿 Drain: $218 (2 jobs)
└ ⚠️ Unclosed Jobs: 1,290
   └ 💰 Hidden Revenue: ~$2.1M
🚨 THE LEAK:
1,290 jobs not marked "completed" in ST. Like having 1,290 invoices stuffed in a drawer that never got deposited.
📏

Confidence Bands = The Honest Estimate

A good plumber never says "It'll cost exactly $4,000." They say "$3,500 to $4,500 depending on what we find when we open the wall." That range is HONEST. It accounts for what you can't see yet.

The Math Engine does the same thing with projections. Instead of "Revenue will be $2.4M this year" (which Stephanie will hold us to and we'll be wrong), we say:
$2.1M
$2.4M
most likely
$2.8M
90% confidence interval based on 13-week trailing data
Now when the actual number comes in at $2.3M, we're not "wrong" -- we're within the range we predicted. The confidence band is the difference between "you lied to me" and "you gave me an honest range."
🔩

Dimensional Analysis = The Wrong Fitting

Every plumber knows: you can't connect a 3/4" copper pipe to a 1/2" PEX fitting. The dimensions don't match. You'll get a leak, or worse, a burst.

The Math Engine applies the same logic to numbers. Every number has a "size" (its unit). Dollars. Leads. Jobs. Percent. You can divide dollars by leads (that gives you $/lead = CPL). But you cannot add dollars to leads. That's like connecting copper to PEX without an adapter -- the math "leaks."
✔️ CORRECT FITTINGS
$6,000 ÷ 130 leads = $46/lead
$208,000 ÷ $6,000 = 34.7x ROAS
23 jobs ÷ 130 leads = 17.7% booking rate
❌ WRONG FITTINGS
$6,000 + 130 leads = ??? (meaningless)
$208,000 ÷ 23 jobs = $9,043/job (not ROAS!)
Revenue - Leads = ??? (unit mismatch)
🧠 THE POINT: When AI calculates ROAS but accidentally divides revenue by JOBS instead of AD SPEND, the number looks legit but it's the wrong fitting. Dimensional analysis catches it. The system REJECTS the calculation before it reaches any report.
💬

Data Storyteller = The Tech's Explanation

A great tech doesn't hand the customer an invoice and walk away. They explain what they found: "Your main line had a 6-foot section of root intrusion from the oak tree. We replaced it with PVC using trenchless method. Here's the before and after camera footage."

The Data Storyteller does the same thing with numbers. Instead of a dashboard showing "$27,949" -- which means nothing without context -- it tells the story behind the number:
"ST shows $27,949 in completed revenue for Q1 -- but that's only 11 of 1,301 jobs actually marked complete (0.8% completion rate). The real revenue is hidden in 1,290 unclosed jobs. QB P&L shows $332,943 for the same period. The $305K gap is techs not clicking Done on their iPads. Top earner: Nick ($19,957, 4 jobs). Top service type: Inspection ($19,500 from one camera job). Action: Fix the completion workflow. Every unclosed job is revenue that can't be tracked, attributed, or proven to a lender."
Three personas, three stories from the same data:
👑 STEPHANIE
"Cross-validate against QB. The $305K gap is the #1 operational problem. Every unclosed job is invisible to lenders and investors."
📞 ASHTON
"Focus on closing the 30 open estimates in Command Center. Nick is producing -- get his estimates to close faster."
🔧 KALEN
"Which techs have the highest zero-invoice rate? That tells you who's doing the work but not closing the paperwork."
🔎

Causal Narratives = The Camera Inspection

You don't dig up a sewer line on a hunch. You run the camera first. The camera shows you EXACTLY where the problem is, what's causing it, and how bad it is. Then you make the call: spot repair or full replacement.

The Causal Validator is the camera inspection for attribution claims. When someone says "Google Ads drove $50K in revenue," the validator runs the camera through the entire pipe:
Ad Click
Landing Page
Phone Call
Paid Invoice
📷 CAMERA FOUND THE BREAK:
The "Paid Invoice" link is broken across ALL attribution chains. payment_collected = $0 for every job. Like finding roots in the main line -- it's blocking EVERYTHING downstream. Fix the completion workflow and the entire attribution system comes back to life.
🌡️

Anomaly Detection = The Pressure Gauge

A plumber doesn't wait for a pipe to burst. They check the pressure gauge. Normal is 40-80 PSI. If it's at 120 PSI, something is wrong BEFORE the damage happens.

The Anomaly Detector watches 6 pressure gauges every hour:
📊
Sync Freshness
"Is new data flowing in?"
Like checking water pressure
📈
Daily Job Count
"Are we booking normally?"
Like checking flow rate
⚠️
Zero Invoice Rate
"Are techs closing jobs?"
Like checking for leaks
API Health
"Are all systems responding?"
Like checking all fixtures
🧠
RAG Integrity
"Is the brain intact?"
Like checking the main shutoff
🌐
Graph Integrity
"Are all connections mapped?"
Like checking pipe routing
When any gauge goes outside the normal range, the system flags it BEFORE anyone notices the damage.
🛡️ Adaptive Immunity: 28 Brain Chunks
8
Error Patterns
Phantom revenue. Dispatch empty. Auto-tagger dry run. Batch edit cascades. Pin gate wrong name. Schedule destroyed. ST 404. Bash heredocs.
7
Prevention Rules
Check job_status. Read before edit. Verify output. Audit batch edits. Test endpoints. Write to file. Query Zeus first.
13
API + Architecture
ST tech ID mapping. Vapi PATCH. QB tokens. 3CX OAuth. WordPress danger. Data flows. Service map. Maturity levels.
Every failure creates an antibody. The system can never make the same mistake twice.
🏰
THE MOAT
The intelligence comes from 5 generations of plumbing in Kansas City.
2,808 RAG chunks. 54 graph nodes. 10 proven formulas. 4 attribution chains.
Graph Brain + RAG Brain + Math Engine + Causal Validator + Context Harness.
No other plumbing company on earth has an AI Operating System.
$7
/month
vs
$50K+
/year as SaaS
Zeus RAG Intelligence Layer | Bright Side Plumbing | (913) 963-1029 | Powered by pgvector + Claude AI