TokenArch
Exploring How Deterministic Systems and LLMs Should Actually Work Together.
I’ve spent 15+ years building the data pipelines that run businesses — SQL, VBA, Python, SSIS, Power BI. Now I’m working through how LLMs fit into that same continuum: where they belong, where they don’t, and what it costs. This site is where I share that thinking as it evolves.
The Landscape
Where Things Actually Stand
There’s a significant gap between what organizations expect from AI and what they’re actually prepared to implement. The data is striking — and it consistently points to the same root causes.
80%
of enterprise AI projects never reach production. Harvard Business Review reports this is nearly double the failure rate of traditional IT projects — and the trend is accelerating, not improving.
70%
of the obstacles come from people and processes, not the technology itself. BCG found that only 10% of AI failures are actually about the algorithm. The rest is organizational: governance, data readiness, architectural clarity.
68%
of enterprises say their data isn’t clean or reliable enough for AI. Meanwhile, 84% of practitioners encounter conflicting versions of the same metric. AI doesn’t solve that gap — it exposes it.
3/4
of organizations lack a clear measure of AI value. Half of CEOs believe their job depends on getting AI right — but without cost visibility at the token level, there’s no way to know if it’s working.
Sources: Harvard Business Review (2023), BCG “From Potential to Profit” (2025), Modern Data Report 2026, S&P Global Enterprise AI Survey (2025), BCG AI Radar via WEF (2026), MIT Project NANDA (2025), Deloitte AI Token Economics (2026).
The Framework
The Numerate Semantic Control Plane
I haven’t seen a great implementation framework for integrating LLMs into enterprise data workflows yet — one that treats them as what they are (word predictors) and keeps everything else deterministic. So I’m building one. The core idea: a semantic layer owns all business definitions and calculation rules, deterministic engines handle computation, and LLM nodes handle only what code cannot — intent classification, exception routing, narrative generation. Every output traces to source data, every action passes through compliance controls, and every inference call is tracked by cost. This is a work in progress. The scenarios below use finance as the example domain, but the five-layer pattern is designed to adapt to any regulated environment.
Disclaimer: All data, schemas, scenarios, and architectures shown here are entirely generic and synthetic. They reference common industry tools and patterns — not the proprietary systems, intellectual property, or internal processes of any current or former employer. Nothing on this site represents or is derived from any organization’s confidential information.
The Problem
For decades, organizations have automated manual work with deterministic code — stored procedures, VBA macros, ETL packages, Python scripts. The logic works, but it’s fragmented: scattered across systems with no single source of truth and no lineage from output back to source. Now organizations are layering probabilistic AI on top of these ungoverned foundations — and the results are predictable. That structural problem predates AI and persists regardless of industry.
The Insight
Much of the confusion I see stems from treating “AI” as a monolithic capability rather than a specific type of node in a larger system. An LLM is a word predictor — powerful at interpreting intent, classifying exceptions, and generating narrative, but fundamentally unsuited for arithmetic, rule enforcement, or compliance logic. Those remain deterministic. The architecture challenge is drawing that boundary clearly — and the business challenge is knowing what each inference call costs.
The Proposed Architecture
Five layers, each with a clear responsibility: source systems feed a governed data fabric; a semantic layer encodes all business definitions and calculation rules; a control fabric enforces compliance (SOX, HIPAA, PCI-DSS); and an agentic layer routes LLM calls through an inference gateway with token-level cost tracking. The blueprint is deterministic — the LLM is invoked only at specific nodes for bounded tasks. The sample schemas below illustrate this with finance data; the pattern adapts to any domain.
NSCP Architecture Stack
5 layers
Layer 5 — Top
Agentic Layer
6 Specialized Agents+Inference Gateway+Token Economics+Budget Controller
Layer 4
Control Fabric
Preventive Controls+Detective Controls+SOX / HIPAA / PCI-DSS+Audit Trail
Layer 3
Semantic Layer
Business Ontology (16 concepts)+Metric Engine (18+ metrics)+Calculation Rules (SQL/Python)
Layer 2
Data Fabric
Raw Zone · CDC · Kafka→Curated Zone · dbt · Quality→8 Governed Data Products
Layer 1 — Bottom
Source Systems
Generic ERPData WarehouseLOB SystemsExcel/VBA/Python Scripts
What This Pattern Delivers
6 capabilities
Numeric Correctness
Agents never perform arithmetic. All calculations execute through deterministic SQL, Python, or rule engines with versioned, testable logic.
Full Auditability
Every KPI traces from presentation → metric definition → calculation rule → source data → control checks. PBC-ready evidence auto-assembled.
Regulatory Compliance
Preventive and detective controls mapped to data products and workflows. Adaptable to SOX, HIPAA, PCI-DSS, or any regulatory framework. No agent takes regulated action or approves its own output.
Token Cost Governance
Inference Gateway logs every LLM call. Budget Controller enforces limits. Model routing optimizes cost. Every AI dollar is allocated and auditable.
Semantic-Agentic Workflows
Agents interpret intent, plan steps, and compose narratives. Deterministic engines run the math. Humans approve material outputs.
Enterprise Integration
Connects ERP, warehouse, LOB systems, SIEMs, or any source system. The semantic layer abstracts sources — agents query concepts, not tables. Swap the domain, and the integration pattern holds.
Live Agent Feed
Simulated
These scenarios use finance workflows as concrete examples — period-end close, AP/AR, bank reconciliation. The underlying architecture pattern (semantic layer → deterministic computation → agent orchestration → compliance controls) applies equally to cybersecurity incident response, healthcare claims processing, supply chain compliance, or any domain where data integrity and auditability are non-negotiable.
Current State — Period-End Close · Pain Points
1
Manual GL Data Extract
Finance analyst logs into ERP, runs ad-hoc SQL query, exports to CSV. No automated scheduling. Data freshness: T+1 day.
Manual
2
Excel/VBA Processing
Analyst pastes data into shared workbook. 2,400-line VBA macro runs transformations. Fragile; breaks on data format changes.
Fragile
3
Email Distribution
Output emailed to 6 stakeholders as attachment. No version control. Recipient edits create conflicting versions.
No Audit Trail
4
Manual Variance Analysis
Senior analyst spends 2–3 hours comparing to prior period. Commentary written in email thread. Not linked to data.
~3 hrs
5
Manual Reconciliation
Cross-system reconciliation done manually in Excel. Breaks require escalation emails. No automated exception tracking.
Error-Prone
6
Journal Entry Posting
Preparer drafts journal in Word, emails to approver. Manual entry into ERP. SOX evidence gathered manually post-hoc.
SOX Risk
Total cycle time: ~4 hours per close · Error rate: ~3.2% requiring rework · Audit evidence: Manually assembled
Target State — NSCP Workflow · Period-End Close · Token Economics
1
Close Planner Agent — Interprets Schedule
Agent reads close calendar, sequences tasks by dependency, assigns agents and SLAs.
1.2K in / 0.8K out~$0.006Claude Sonnet 4.6
Agent
2
Semantic Layer — Metric Definition Lookup
Deterministic resolution of metric IDs to SQL definitions from the metric catalog. No inference required.
0 tokens (deterministic)
Deterministic
3
SQL Execution Against GL Event Store
Compiled SQL runs against fact_gl_event. Results verified against pre-defined control totals. Audit log written.
0 tokens (deterministic)
Deterministic
4
Variance Analysis Agent — Explains Movements
Receives structured variance data. Generates natural-language explanations grounded in semantic-layer definitions.
3.1K in / 1.5K out~$0.014Claude Sonnet 4.6
Agent
5
Control Validation — Automated Rules Engine
16 SOX controls evaluated deterministically. Pass/fail written to fact_control_execution.
0 tokens (deterministic)
Controls
6
Exception Triage Agent — Reviews Flags
Classifies 3 flagged exceptions by severity. Routes 2 to auto-resolve, 1 escalated to Controller.
2.8K in / 1.2K out~$0.012GPT-5.2
Agent
7
Compliance Agent — Drafts Journal Memo
Generates SOX-compliant evidence package: journal entries with rationale, control results, approver routing.
1.9K in / 2.1K out~$0.018Claude Sonnet 4.6
Agent
Note: Token inference cost only. Total cost of ownership includes platform infrastructure (cloud compute, data pipelines, storage), engineering effort to build and maintain, and ongoing operational overhead. This comparison illustrates the marginal cost of AI inference vs. equivalent human labor — not full project ROI.
Current State — AP/AR Cash Cycle · Pain Points
1
Invoice Receipt & Manual Entry
AP staff manually keys invoice data into ERP. OCR tools used inconsistently. Duplicate entry risk.
Manual
2
3-Way Match — Manual
PO, receipt, and invoice matched manually in spreadsheet. 8% exception rate requires supervisor review.
8% Exceptions
3
Aging Report — Emailed Weekly
AR aging exported from ERP, formatted in Excel, emailed to collections team. Data is 24h stale.
Stale Data
4
Cash Application — Manual
Analyst matches bank deposits to AR invoices. Partial payments split manually. High error rate on short payments.
Error-Prone
Target State — NSCP · AP/AR Cash Cycle · Token Economics
1
Automated Invoice Ingestion (OCR + Rules)
Structured extraction from PDF invoices. Rules-based field validation against vendor master.
0 tokens (deterministic OCR pipeline)
Deterministic
2
Automated 3-Way Match
SQL join across fact_ap_invoice, PO table, GR/GI records. Automated matching with threshold tolerances.
0 tokens (deterministic)
Deterministic
3
Exception Triage Agent — AP Exceptions
Reviews unmatched invoices. Classifies root cause. Suggests resolution action.
2.1K in / 1.0K out~$0.009GPT-5 Mini
Agent
4
Automated Cash Application
Bank feed matched to AR invoices using multi-criteria scoring. Auto-applied for 94% of payments.
0 tokens (deterministic)
Deterministic
5
AR Analyst Agent — Aging Commentary
Generates structured aging commentary, identifies at-risk accounts, recommends collection actions.
1.8K in / 0.9K out~$0.007Claude Haiku 4.5
Agent
Note: Token inference cost only — excludes platform infrastructure, engineering, and maintenance overhead.
Current State — Bank Reconciliation · Pain Points
1
Bank Statement Download
Analyst logs into online banking, downloads MT940 or CSV statement. Reformatted in Excel for import.
Manual
2
GL Extract & Comparison
GL cash balance extracted separately. VLOOKUP-based matching across two worksheets. ~15% unmatched on first pass.
15% Unmatched
3
Break Investigation
Analyst researches each break individually. Timing differences documented manually. Escalations via email.
~2hr Investigation
Target State — NSCP · Bank Reconciliation · Token Economics
1
Automated Bank Feed Ingestion
Direct bank API integration. MT940 parser normalizes transactions into fact_recon_match staging.
0 tokens (deterministic)
Deterministic
2
Automated Transaction Matching
Multi-pass SQL matching: exact → amount/date tolerance → fuzzy reference. Auto-clears 97.3%.
0 tokens (deterministic)
Deterministic
3
Reconciliation Agent — Break Analysis
Analyzes remaining 2.7% unmatched. Cross-references payment descriptions, vendor history. Proposes resolutions.
2.4K in / 1.1K out~$0.010Claude Sonnet 4.6
Agent
4
Compliance Agent — Recon Sign-Off Package
Compiles reconciliation evidence: matched summary, break notes, approver routing.
1.5K in / 1.8K out~$0.011Claude Sonnet 4.6
Agent
Note: Token inference cost only — excludes platform infrastructure, engineering, and maintenance overhead.
8 Sample Data Products
GL Event Store
fact_gl_event
AP Invoice Mart
fact_ap_invoice · dim_vendor
AR Invoice Mart
fact_ar_invoice · dim_customer
Recon Match Log
fact_recon_match
Asset Register
fact_asset · dim_asset_class
Control Execution Log
fact_control_execution
Agent Action Log
fact_agent_action
Token Usage Telemetry
fact_token_usage
Sample Business Ontology — 16 Entities
Metric Catalog
| Metric ID | Name | Category | Formula / SQL | Data Products | Target |
|---|
Regulatory Controls & Token Governance
| ID | Control Name | Type | Frequency | Owner | Data Products | Status |
|---|
Tools & Analysis
Token Economics
If AI costs are consumption-based and unpredictable, organizations need new tools to think through them. This is an evolving toolkit for exactly that: model pricing comparisons, inference gateway architecture, ROI modeling, and optimization strategies — grounded in current provider rates and real cost structures. Deloitte calls tokens “the new currency of AI.” This section explores what that means in practice.
Disclaimer: All pricing, calculations, and architectures are generic and synthetic. They reference publicly available provider rates and common infrastructure patterns — not the proprietary systems or intellectual property of any current or former employer.
Representative pricing as of February 2026. LLM API rates change frequently — verify with providers before use in production cost models. Sources: provider pricing pages.
| Model | Provider | Input $/1M | Output $/1M | Context | Tier | Best For |
|---|---|---|---|---|---|---|
| Claude Sonnet 4.6 | Anthropic | $3.00 | $15.00 | 200K | 1 | Financial analysis, compliance docs |
| GPT-5.2 | OpenAI | $1.75 | $14.00 | 200K | 1 | Complex reasoning, exception triage |
| GPT-5 Mini | OpenAI | $0.25 | $2.00 | 200K | 2 | Forecasting, high-volume tasks |
| Claude Haiku 4.5 | Anthropic | $1.00 | $5.00 | 200K | 2 | Budget checks, routing |
| Gemini 2.5 Flash | $0.30 | $2.50 | 1M | 2 | High-volume ad-hoc queries | |
| DeepSeek V3.2 | DeepSeek | $0.28 | $0.42 | 128K | 2 | Classification, extraction |
| GPT-5 Nano | OpenAI | $0.05 | $0.40 | 400K | 2 | High-volume simple tasks |
Tier 1 vs Tier 2 Savings
75–95%
Routing from Claude Sonnet 4.6 to DeepSeek V3.2 or GPT-5 Nano yields up to 95% cost reduction per call.
NSCP Model Policy
Tier 1SOX reporting, compliance, complex analysis
Tier 2Ad-hoc queries, dev/test, classification
BudgetAuto-downgrade at 80% utilization
Blended Cost/1K Tokens
80/20 input/output token split
Claude Sonnet 4.6$0.0054
GPT-5.2$0.0042
GPT-5 Mini$0.0006
DeepSeek V3.2$0.0003
Inference Gateway Architecture
Token Tracking PipelineForward Path — Request
Agent
Initiates call
Initiates call
→
Inference
Gateway
Auth + Route
Gateway
Auth + Route
→
Pre-Call Check
Budget + Token Est.
Budget + Token Est.
→
Model Router
Tier enforcement
Tier enforcement
→
LLM API
Provider
Provider
Return Path — Response + Logging
Agent
Receives result
Receives result
←
Gateway
Strip metadata
Strip metadata
←
Post-Call Log
Actuals + Budget
Actuals + Budget
←
LLM Response
+ Usage headers
+ Usage headers
Downstream Consumers
Token Logger
→
fact_token_usage
→
Token Economics Dashboard
Budget Controller Agent
TGC Controls (TGC-001–006)
Gateway Design Targets
Overhead: <12ms p99
Availability: 99.95%
Peak RPS: 247
Cache hit: 44.2%
Availability: 99.95%
Peak RPS: 247
Cache hit: 44.2%
Pre-Call Checks
✓ Agent budget available?
✓ Model tier authorized?
✓ Token estimate within limit?
✓ Prompt hash — cache check
✓ Rate limit headroom?
Post-Call Actions
→ Write to fact_token_usage
→ Decrement agent budget
→ Update cache registry
→ Fire alerts if threshold crossed
→ Tag cost to business unit
Calculator Inputs
Input/Output split: 80% input / 20% output tokens
Payback assumes estimated platform setup cost ($5K–$60K based on volume)
Payback assumes estimated platform setup cost ($5K–$60K based on volume)
Monthly AI Cost
—
Monthly Human Cost
—
Monthly Savings
—
ROI
—
Cost/Run — AI
—
Cost/Run — Human
—
Value per Token
—
Payback Period
—
Prompt Caching
Cache frequently used system prompts and static context. Dramatically reduces input token processing cost.
Savings: 50–90% on cached portions
Intelligent Model Routing
Route simple tasks to Tier-2 models. Reserve Tier-1 for complex reasoning, SOX compliance, and financial judgment.
Savings: 60–80% on routed calls
Batch Processing
Group similar requests to amortize system prompt overhead. Effective for reconciliation checks and exception classification.
Savings: 20–30% on batch-eligible tasks
Token Budget Allocation
Set daily/monthly token budgets per agent. Controller enforces hard limits, triggers downgrades at 80%, blocks at 90%.
Prevents runaway costs
Prompt Engineering
Optimize prompt structure to reduce tokens without degrading quality. Use structured output formats.
Typical reduction: 15–25% on input tokens
Semantic Caching
Cache responses for semantically similar queries using embedding-based similarity. Return cached responses with metadata.
Hit rate improvement: 10–20%
Combined Optimization Impact
| Optimization | Applicable Calls | Reduction | Monthly Savings | Effort |
|---|---|---|---|---|
| Model Routing | 70% | 65% | $812 | Low |
| Prompt Caching | 44% hit | 70% | $218 | Med |
| Prompt Engineering | 100% | 20% | $110 | Med |
| Batch Processing | 35% | 25% | $64 | High |
| Semantic Caching | 15% | 100% | $47 | High |
| Total (combined) | ~80% | $1,251/mo | ROI: 12x |
Background
The Path Here
From early web development through financial domain depth to enterprise systems leadership — I’ve built my career at the convergence of finance, data, and technology.
Web Development & Technical Foundations
Built websites and web applications for small businesses — first professional exposure to systems thinking, client delivery, and working code in production.
Associate's — Information Systems
Formal CS foundation: databases, networking, software development, systems analysis.
Bachelor's — Finance
Bridging business and technology. Financial theory, accounting principles, corporate finance.
Wharton — Business & Financial Modeling
Wharton-trained in advanced financial modeling, valuation frameworks, and strategic quantitative analysis.
Oracle Cloud Certification
Enterprise platform expertise: ERP architecture, cloud infrastructure, integration patterns.
Valuations Coordinator
Financial domain depth: valuation methods, market analysis, model construction.
Valuations Auditor
Controls, compliance, and risk mindset. SOX testing, evidence standards, audit methodology.
Business Data Analyst
Analytics and reporting: SQL, VBA, Python, SSIS, SSRS, data visualization, stakeholder communication, KPI frameworks.
Sr. Financial Systems Analyst
Enterprise systems ownership: ERP administration, integration architecture, process automation.
Lead Finance BI Developer
Data platforms and leadership: team management, roadmap ownership, cross-functional delivery.
This background means I speak finance and engineering fluently — the rare overlap that lets me diagnose in days what keeps departments stuck for quarters.
Fortune 500 to Startup — Proven at Every Scale
Wharton-Trained — Business & Financial Modeling
Oracle Cloud Certified
Years Cross-Functional Experience
What I Work On
Finance Data Architecture
Governed, auditable data architectures — from ERP integration through warehouse design to compliance-ready pipelines.
- Enterprise GL, AP, AR, and reconciliation data platforms
- ERP integration, ETL pipelines, data warehouse design
- Data governance, quality frameworks, SOX-compliant architectures
SQL
Python
Oracle Financials
SOAP/REST
XML Publisher
BI & Analytics Engineering
Dashboards, reporting, and analytics platforms that turn raw data into decisions — built by someone who understands the underlying business logic.
- Executive dashboards, self-service analytics, operational reporting
- Power BI, Tableau, SSRS development and optimization
- KPI frameworks, metric catalogs, semantic layers
Power BI
Tableau
SSRS
SSIS
VBA
Python
AI-Ready Architecture & Governance
Forward-looking architecture thinking for how LLMs integrate into enterprise workflows — with the governance frameworks and cost visibility that production deployment will require.
- Agentic workflow design — LLMs as semantic routers, not calculators
- Token economics and inference cost tracking — the metrics executives will need before approving any AI workflow
- The Numerate Semantic Control Plane — a reference architecture demonstrated below
- Governance-first approach: audit trails, human approvals, regulatory-compatible controls
Python
LLM Architecture
AI Governance Design
Architecture Design
Contact
Let’s Talk
I’m sharing these frameworks and tools because I think the conversation around enterprise AI needs more architectural rigor and less hype. If any of this resonates — or if you see gaps I should address — I’d welcome the conversation.