The platform a bench technician would have built. If they'd had an AI team.
102 Features4 Phases3.06× Delivery LeverageMVP: Q1 2027On-Prem First
01 - The Opportunity
The gap between institutional knowledge and scalable throughput
The challenge today
Expertise lives in heads, not systems
Senior technicians carry diagnostic knowledge that is not captured anywhere
Bench diagnosis time is long because each tech starts from scratch
Four sites, no shared knowledge layer - the same problem is solved four times
Seasonal peaks overload intake and stretch turnaround windows past customer tolerance
No closed loop between what was diagnosed and whether the fix actually held
The platform opportunity
An agentic knowledge layer built on your data
Capture every technician insight as a structured graph entry - permanently
Guided bench diagnosis from a corpus of real cases, traces, and resolved failures
One shared intelligence layer across all four sites from day one
Voice intake agents that handle call volume during peak season without headcount
Outcomes feed back into the model: every repair makes the next diagnosis sharper
The gap between a $400 automotive signal simulator and a $250k OEM validation rig is the exact space Ag Express has occupied for 34 years. The knowledge to close that gap already exists - in repair tickets, harness catalogs, and 34 years of multi-brand failure data. The platform makes it queryable, shareable, and self-improving.
02 - The Platform
Five nodes. One intelligence layer.
Each node is a governed business capability with its own data contract and agent scope. They share the knowledge graph and accumulate value together over time.
Live data flows · teal: platform nodes · amber: in-house manufacturing loop (first-party data advantage) · blue: OEM/partner nodes · pulse ring: seasonal intake peak
Command Center
On-prem control plane with local model serving, knowledge graph, vector index, and full data sovereignty. The foundation everything else runs on.
Foundation / Phase 1
Corpus
Structured ingestion of repair tickets, harness catalogs, OEM manuals, failure mode taxonomies, and technician expertise into a queryable graph.
Knowledge / Phase 1–2
Bench
CAN/J1939 hardware interface, trace replay, response differencing, guided diagnostic flow, verdict with confidence, and quote generation.
Diagnostic Core / Phase 1–2
Intake
Voice agents handling inbound repair calls with transcription, intent routing, ticket creation, status answers, and warm human handoff.
Voice / Phase 2
Operations
Repair queue routing across four sites by capacity and capability, turnaround forecasting, parts demand signal, and ISO evidence generation.
Ops / Phase 2–3
03 - The Moat
Ag Express manufactures what it repairs. No competitor can replicate that.
Ag Express operates an in-house harness manufacturing division certified under ISO 9001:2015, covering wire harness manufacturing, box builds, and electronics assembly. The harnesses that division builds are the same harnesses that arrive on the Ag Express bench.
The data advantage
First-party specs on first-party failures
For harnesses and products manufactured in-house, Ag Express doesn't need OEM permission to build the trace reference library. You are the OEM. Connector pinouts, expected signal ranges, failure modes - all first-party data with full IP rights. This cannot be acquired, negotiated, or replicated by a competitor.
What this enables
The strongest possible corpus baseline
Trace replay: in-house harness traces are the cleanest possible reference - no OEM interpretation, no reverse engineering, no field-of-use restriction
Phase 4 data product: in-house failure data has inherently higher quality than any third-party OEM harness data
Start the trace library with Vanguard and Ag Express-manufactured devices for zero-restriction knowledge entries
Manufacturing customers are natural first dealer tenants - existing relationships, existing trust
Right-to-repair positioning
The July 2026 FTC/Deere settlement requires Deere to give independent repairers the same software access as dealers. [source] Ag Express has something stronger: on products it manufactured, it has data rights that no dealer and no OEM will ever grant a competitor. This data advantage is not an edge - it is a wall.
04 - The Substrate
A seven-layer cognitive architecture. Each layer activates as the engagement matures.
The platform doesn't arrive as a finished product. It starts as a proven foundation and compounds. Each layer makes the next one possible. By Phase 4, the system is self-auditing and self-improving on your data - on your hardware.
L2
Memory Phase 1
Neo4j knowledge graph with three-tier memory: episodic (every bench session, time-weighted and retrievable), semantic (all failure modes, queryable in milliseconds), procedural (approved diagnostic steps, technician-verified before going live). 34 years of institutional knowledge stops living exclusively in senior technicians' heads.
L4
Routing Phase 1
Model routing, cost management, and failover. 95% of routine lookups stay on-prem and return in under 100ms. Cloud models handle only novel or high-complexity cases. No per-token costs on your own corpus - the hardware pays for itself within months and the cost curve inverts from there.
L1
Values Phase 2
The bench agent's governance layer, built before verdict output goes live: never produce a diagnosis without citing a corpus entry; flag uncertainty above a set threshold; escalate to a human technician when outside training distribution. The liability guardrail is structural, not advisory - it cannot be overridden by a poorly-worded prompt.
L3
Perception Phase 2–3
Event ingestion bus: every bench signal, technician action, corpus update, and state change flows through a single audit trail. ISO 9001 evidence generation is a byproduct of normal operation, not a separate feature built at extra cost.
L0 / L5
Prediction + Audit Phase 3–4
Predictive failure patterns built from accumulated corpus data. The audit layer (L5) self-checks: if a diagnostic pattern the system rates highly keeps getting corrected by technicians, it flags itself for review before reaching the queue again. Confidence calibrates to evidence, not assumption.
L6
Self-learning Phase 4
Every technician-approved bench session is a training example. When approved examples cross threshold, the system fine-tunes on your data - on your hardware. After Phase 4, you have an agricultural electronics diagnostic model that no competitor can replicate because no competitor has your 34-year corpus.
How the substrate is delivered
The platform runs on your hardware under a managed subscription. We operate it, push improvements as it matures, and maintain it continuously. Your data is always on your hardware - never in our systems. After 12 months of live operation, you have the option to convert to a perpetual license and own the stack outright. That conversion price reflects what you helped build. Anonymized failure signal patterns from your corpus also help improve the platform itself, which means your stack benefits from every platform-wide improvement without paying separately for R&D.
05 - Why Swift Innovation
Built for this kind of work
Full-stack software and agentic systems
We design and build client-facing software products across web, systems, and agentic domains. The portal you are reading this on - with its access control, document delivery, and live feedback layer - is representative of what we ship, not what we're planning to build.
On-prem first, cloud optional
Your data and models stay on your hardware. We have direct experience deploying local inference servers, graph databases, and CAN interface hardware in industrial environments.
Domain knowledge compounds
Every bench session, every captured technician insight, every repaired unit goes back into the corpus. The system gets materially better every month. That is not an AI marketing claim - it is the design.
Agentic build methods compress delivery time without compressing quality. Multiple tracks run simultaneously instead of each phase waiting for the last.
Traditional sequential build
Sequential build. Every phase waits for the last.
Infrastructure setup
↓ waits
Data ingestion & corpus
↓ waits
Bench application
↓ waits
Voice intake
↓ waits
Operations layer
Agentic parallel build (Program tier)
Parallel build. Infrastructure, knowledge layer, and bench run together.
Command Center
→
Bench MVP
Corpus ingest
→
Knowledge graph
Intake voice agents
→
Ops queue
✓ MVP live: end of Q1 2027
3.06×
delivery leverage
1,521
agent hours
4,658
human-equivalent hours
102
features, 4 phases
AI agent coverage - active agents at each phase
Capability
Node
Agent type
Phase
Knowledge graph population
Corpus
Ingestion agent
1
Guided bench diagnostic flow
Bench
Diagnostic agent
2
Verdict and repair recommendation
Bench
Verdict agent
2
Inbound call intake and routing
Intake
Voice agent
2
Repair status answering
Intake
Voice agent
3
Queue routing across sites
Operations
Ops agent
2
Corpus quality and retraining
Command Center
Quality loop agent
3
07 - Ownership
Your data, your application layer. Always.
There are two layers to the platform. The application layer (bench tool, corpus, work order system, agent configurations) transfers fully to Ag Express at each phase milestone. The platform substrate (the cognitive infrastructure underneath) runs as a managed service on your hardware under the retainer.
✓
All your data - always on your hardwareKnowledge graph, corpus, repair history, training examples, technician knowledge - none of it ever leaves your infrastructure
✓
Bench application source codeFull desktop application codebase delivered at the end of Phase 2, version-controlled and documented
✓
Agent configurations and governance rulesAll agent charters, prompts, and approval rules are version-controlled and exportable at any time
✓
Continuity package (Phase 2)Runbooks, escrow, and rehearsed handover delivered before Phase 3 begins (CC-18)
✓
Fine-tuned domain model (Phase 4)Domain-adapted model trained on your corpus on your hardware. Weights owned by Ag Express; eval harness included
✓
No metered API costs on your own dataOn-prem inference means no per-token billing on your own corpus queries. Local hardware, fixed cost
✓
Platform substrate: managed subscriptionThe cognitive infrastructure runs as a managed service on your hardware. We operate it; you own the hardware and the data
✓
Perpetual license path at 12 monthsAfter the first year of live operation, you have the option to convert the platform substrate to a perpetual license. That price reflects what you helped build and mature
The application layer (bench tool, corpus, work order system, all custom code) transfers to Ag Express continuously. The platform substrate is operated for you under the managed subscription - think of it as the OS that runs under the application, maintained and updated by us on your hardware.
You get what we run in production - installed on your hardware, owned permanently. As the platform improves, upgrades are your choice on your timeline. You are never starting from scratch.
08 - The Plan
Phase 0 then four build phases. MVP end of Q1 2027.
Phase 0 is a fixed-fee pre-build engagement: hardware is identified, procured, and installed before a single line of product code is written. Phases 1-4 timelines shown for the recommended Program tier (3 engineers, parallel build). Foundation and Build tiers extend each phase proportionally.
Phase 0
Discovery
Sept 2026 - 2-4 wks
Phase 1
Foundation
Oct 2026 - 1 mo
Phase 2 - MVP
Core Systems
Nov 2026 - Feb 2027
Phase 3
Scale
Q2-Q3 2027
Phase 4
Platform
Q4 2027
Hardware live
Oct 2026
Corpus seeded
Nov 2026
Bench MVP live
Q1 2027
Data product
Q4 2027
Phase 0
Discovery + Infrastructure
Sept 2026 - 2-4 weeks
Engagement signed: NDA, agreements, site access
Hardware specification: GPU inference server, NAS, networking scoped to workload
Procurement: all hardware purchased in Ag Express name, at cost, no markup
Rack, power, and network installation at primary site
Connectivity and baseline performance tests
Phase 1 scope confirmed, kickoff scheduled
Fixed fee · 2-4 weeks Phase 1 starts with hardware live
Phase 1
Foundation
Oct 2026 - 1 month
On-prem control plane
Local model serving + router
Knowledge graph + vector index
Ag electronics ontology
Ingestion pipeline + classification
Baseline measurement
13 features · 169 agent hrs
Substrate: L2 Memory + L4 Routing live
Phase 2 - MVP
Core Systems
Nov 2026 - Feb 2027
Bench hardware interface + desktop app
Trace replay + response differencing
Guided diagnostic + verdict
Corpus: technician knowledge + trace lib
Voice intake + call routing
Repair queue orchestration
40 features · 495 agent hrs
Substrate: L1 Values + L3 Perception added
Phase 3
Scale
Q2–Q3 2027 - ~4 months
Field tablet application + offline mode
Behavioral model runtime
Advanced voice capabilities
Web + marketing content pipeline
Parts demand signal
Corpus quality loop
32 features · 421 agent hrs
Substrate: L0 Prediction + L5 Audit added
Phase 4
Platform
Q4 2027 - ~4 months
Domain model fine-tune
Component-level fault localization
Dealer tenancy + licensing
Electronics residual risk scoring
External data product distribution
Internal knowledge assistant
13 features · 436 agent hrs
Substrate: L6 Self-learning active
09 - The Team
Right-sized allocation at every tier
Each retainer tier includes dedicated engineering capacity plus shared leadership. The agent layer is active across all tiers and multiplies throughput without adding headcount.
Role
Foundation
Build
Program
Senior Engineers
1
2
3
Architecture & PM leadership
✓
✓
✓
On-prem hardware support
✓
✓
✓
AI dev tooling overlay
✓
✓
✓
Leadership time (architecture, PM, and QA) is shared across clients but fully dedicated to Ag Express during active sprint planning, review, and handoff. Engineers are assigned full-time to this engagement per tier.
10 - Investment
Three tiers. One objective: MVP by end of Q1 2027.
All tiers reach the same platform. The tier determines the pace and team size. The recommended tier - Program - targets MVP delivery within the winter trough window when operational disruption is lowest.
Foundation
$10k/mo
1 engineer + leadership
Phase 1 Foundation (~3 months)$30k
Phase 2 Core (~9 months)$90k
MVP timeline~12 months
Phase 3 Scale (~9 months)$90k
Phase 4 Platform (~9 months)$90k
One-time onboarding fee$10k
Total through MVP$130k
Best for a measured start. MVP delivery by ~Oct 2027.
Build
$20k/mo
2 engineers + leadership
Phase 1 Foundation (~2 months)$40k
Phase 2 Core (~5 months)$100k
MVP timeline~7 months
Phase 3 Scale (~5 months)$100k
Phase 4 Platform (~5 months)$100k
One-time onboarding fee$10k
Total through MVP$150k
Bench and knowledge layer built concurrently.
Program - Recommended
$30k/mo
3 engineers + leadership
Phase 1 Foundation (1 month)$30k
Phase 2 Core (4 months)$120k
MVP timeline5 months
Phase 3 Scale (4 months)$120k
Phase 4 Platform (4 months)$120k
One-time onboarding fee$10k
Total through MVP$160k
Winter trough build window. MVP live end of Q1 2027.
11 - Next Steps
Four steps to a signed engagement
01
Discovery call
Review the proposal and feature matrix together. Walk through Phase 1 and 2 scope. Confirm hardware inventory and site access.
02
Agreement and tier selection
Select the retainer tier. Sign the services agreement. The agreement includes IP assignment, continuity provisions, and the data classification policy.
03
Infrastructure provisioning
Deploy the on-prem control plane and establish the knowledge graph. Run the baseline measurement (CC-17) to instrument current performance before anything changes.
04
Phase 1 kickoff
Sprint 1 begins. Corpus ingestion starts on the first available repair ticket dataset. Command Center goes live at the end of the first month.
The on-prem design eliminates recurring AI compute costs. The choice is a one-time capital outlay versus paying cloud API rates every month, per engineer, indefinitely.
On-prem hardware (one-time capital cost)
Equipment
Notes
Est. one-time
GPU inference server
24–48 GB VRAM; runs all primary models locally
$8,000–$12,000
NAS / bulk storage
Corpus ingestion, vector snapshots, backups
$1,000–$2,000
Managed network switch
VLAN isolation for inference traffic
$500–$1,000
Cabling and UPS
Rack power protection and cable management
$300–$600
Hardware total
Purchased at Phase 1 kickoff, in Ag Express's name, at cost - no markup
~$10,000–$15,000
Alternative: cloud API tokens (per engineer, per month)
Cloud API route
Pay-as-you-go tokens
Per engineer / month$1,000–$1,500
Foundation (1 eng)$1,000–$1,500/mo
Build (2 engs)$2,000–$3,000/mo
Program (3 engs)$3,000–$4,500/mo
Cost is ongoing and scales directly with team size and usage. No ceiling.
On-prem route (recommended)
Own the compute
Capital cost (all tiers)~$10k–$15k once
Recurring inference cost$0/mo
Payback vs. Foundation8–15 months
Payback vs. Program3–5 months
Hardware pays for itself before MVP launches at the Build or Program tier. Every month after that is pure savings.
Post-launch steady-state operating costs
Line item
Notes
Est. monthly
On-prem compute (inference)
Ag Express hardware - no cloud bill
$0
Knowledge graph database
Neo4j Community or equivalent on-prem
$0–$150
Vector store
Weaviate or Chroma self-hosted
$0–$100
Voice / telephony (Intake)
Per-minute billing; scales with call volume
$200–$600
Backup and offsite storage
Encrypted snapshots; object storage
$50–$150
Frontier model API (overflow)
Optional; local models handle primary load
$100–$300
Monitoring and alerting
Self-hosted Grafana + Prometheus
$0–$50
Post-launch total estimate
~$350–$1,350/mo
Voice costs vary with inbound call volume and season. The on-prem model serving eliminates per-token API costs that would represent the largest line item in a cloud-first design. Infrastructure estimates are reviewed at the end of Phase 1 once actual call volumes and workloads are instrumented.