Architecture & Security Model

How It Actually Works

A custom-built system. Knowledge hubs, workflow tools, and one shared platform over the institutional knowledge your existing systems cannot answer from. This page is the two questions every technical reviewer asks, answered in full.

Question 01

"We already have a CRM, an ERP, and a document system. Isn't this just a shell over data we already have?"

Question 02

"What happens to our information, especially once it touches our systems of record, and how is it protected?"

Status: Proven in a reference deployment
Scope: Built custom per engagement
Read time: 12 minutes
Section 01

The category error. "Isn't this just a shell?"

What most teams have today

Several calculators, a document store of reference PDFs, a system of record, a vendor list, a directory. Each answers its own narrow question, as long as you already know which one to open:

What gets built

The expensive knowledge is different in kind. Audit expectations, customer rules, trade-offs, and who to ask live in standards, transcripts, and senior heads. No system you own can answer:

Not a replacement. Your ERP stays the system of record. Your document store stays the document store. The platform reads from and writes to them only where the engagement scopes it. Nothing is migrated, nothing is duplicated.

Not a wrapper. It does the work between the systems: the interpretation, the cross-referencing, and the "what does this choice cost me." The hours that currently happen inside one person's head.


Section 02

Two kinds of module, one platform

Every build is assembled from two module types standing on one shared platform. Modules marked proven exist in production today. Modules marked custom are designed and priced per engagement.

Knowledge hubs: grounded Q&A agents, one machine, your corpora
Proven

Compliance / Audit Helper

Chat, tier self-assessment with a printable gap report, and a browsable rubric.

Proven

Regulatory Q&A

Chat, a scope wizard, and a digitized inspection checklist with escalation cards.

Custom

Customer Requirements Hub

Per-customer specs, approved options, labeling and delivery rules.

Custom

Policy Guidelines Agent

Sustainability, cost, or risk policy applied to a live decision.

Custom

Technology Education Hub

Options, trade-offs, and when each one fits, for the decision being made.

Custom

Inventory / Surplus Slice

What you already own, before anyone buys new.

Workflow tools: structured state, connected
Custom

Unified Calculator

Replaces N spreadsheets with one chat-and-form hybrid: editable assumptions, help bubbles, compliance prompts at the right step.

Custom

Vendor / Supplier Directory

Smart filtering by capability, region, and approval status.

Custom

Contacts + Teams / Slack

Routes any unknown to the named owner, with context already attached.

Custom

Form to System of Record

Calculator output auto-populates the planning form, which feeds your ERP after a human confirms.

The shared platform: every module stands on this

Gate and your SSO. Versioned knowledge store with an owner-operated update pipeline. Grounded-agent framework (knowledge plus rules layers). Cross-link bus, the criteria tags that fire prompts across modules. Sliced training-video help bubbles. Audit log. Integration layer, built in the last phase, inside your tenant.


Section 03

Not just a chatbot. One knowledge base, three ways in

Ask it

Grounded chat. Cited answers in plain language, and a refusal instead of a guess when the question falls outside the loaded corpus. "How do we get to the next tier?" "What do I do on a failed inspection?"

Check against it

A self-assessment walks your scoring ladder and produces a printable gap report: current tier, the red-line warning, and the reviewer's evidence list for every gap. Scope wizards answer "does this rule apply to me?" Digitized checklists end in your SOP's exact escalation card on any failure.

Read it

The entire corpus as a browsable, deep-linkable reference. Every criterion with its requirement, evidence checklist, and practical guidance. It updates the moment your owner publishes a new corpus version.

The chat gets the demo applause. The structured tools are what people open on a Tuesday, and the deep links are how the knowledge spreads through the organization on its own.


Section 04

Named parts, no magic boxes

The request path, for every module and every user request:

01 Client

Browser

Gated session, httpOnly cookie, no secrets client side.

02 TLS

Server application

Managed cloud. Assembles the grounded prompt or calculator state and streams back.

03 TLS

Frontier language model

Commercial API, under contractual no-training terms.

04 Reads

Knowledge store

Versioned corpora. No personal data, by design.

Standard parts, deliberate assembly. The value is the knowledge layer, the cross-link bus, and their governance. Built once, reused by every module in your build.


Section 05

How it all works together. One person, one decision

1

Scope

Inputs, volumes, candidate options. The first-pass estimate. The Customer Requirements Hub injects that customer's approved options and rules, the Vendor Directory offers suppliers who can actually deliver the candidate, and the inventory slice checks what you already own first.

2

Size

The quantitative core: fleet, capacity, cycle, safety stock. A compliance prompt fires with the criterion this step must satisfy and the evidence it needs. The technology hub offers add-on options and their compliance implications. A help bubble plays the 40-second training slice for the field being edited.

3

Optimize

Trade-offs: cost against policy against risk, and right-sizing. The policy agent runs the comparison your guidelines require and shows the exposure. A compliance prompt says to log this as the documented improvement. A regulatory check runs if anything regulated enters the spec.

4

Plan

Calculator output auto-populates the planning form. A human reviews and confirms every field before anything is written. Contacts routes any open question to the named owner with context attached.

5

Record

The confirmed form feeds your system of record, which stays the system of record. The integration layer uses a least-privilege service account, writes only after confirm, and logs every write. Runs inside your tenant, coordinated with your IT, in the final phase.

Today this is several spreadsheets, a document store, and a phone call. In the platform the person moves left to right and the knowledge shows up at the step where it matters: the customer rule at Scope, the compliance consequence at Size, the policy math at Optimize. Instead of being remembered, searched for, or missed.

The step names above are the reference deployment's. Yours are mapped in the first scoping session.


Section 06

The cross-link bus. How one module fires inside another

Every criterion in a knowledge hub carries structured tags naming the workflow moments it governs. When a tool reaches one of those moments, it asks the platform "what applies here?" and the relevant hub answers, in context, citing its source.

Nothing is hard-coded between modules. Add a criterion, tag it, and it starts firing in the right place. Retire it, and it stops. It is the difference between nine tools and one system, and you can feel it: in the reference deployment, the calculator preview pops the real compliance prompt the moment a tracked step is toggled off. Sample data, real wiring.

When a user unchecks a tracked or required step in the calculator
Compliance prompt: "This drops you below the tier. Here is what the reviewer will ask for, and the options that satisfy it."
Compliance helper → Unified calculator → Technology hub
When a spec includes a regulated material, process, or document
Regulatory prompt at the planning form: the certification, inspection, or sign-off required, before the form can be confirmed.
Regulatory Q&A → Planning form
When a lower-cost option is chosen where policy prefers another
Policy agent shows the required comparison, and the matching compliance evidence requirement attaches to the decision.
Policy agent → Calculator → Compliance helper
When any module cannot answer with confidence
It refuses to guess. Contacts routes the question to the named owner in Teams or Slack, with the context attached.
Every hub → Contacts

Tag schema and resolution logic are licensed IP.


Section 07

Six hubs, one machine. Why they do not hallucinate

Layer 1: knowledge, one per hub

Each corpus is versioned, curated with its owner, and injected server side. Never editable by the person asking.

Layer 2: rules, shared by all hubs

Binds the model to its corpus: never invent, refuse outside loaded scope, quote the source verbatim where it matters, route unknowns to the named owner. Same rules layer every time, so a hub that passes the test once passes it for every corpus. Construction of both layers is licensed IP. The corpus is cached at the API level, so it is billed once per cache window rather than once per question.

Tested adversarially, not assumed

In the reference deployment, an internal audit-prep hub for a Fortune 500 Tier 1 automotive supplier, 24 realistic questions were fired at the production endpoint before its first internal demo. Three were probes designed to bait fabrication:


Section 08

The part no off-the-shelf tool has. The knowledge pipeline

Standards get rewritten. Customers revise their specs whenever they like. Most AI tools would need a rebuild. Here, each hub's named owner updates its corpus through one pipeline serving every hub. No code change, no redeploy, no IT ticket. This is what makes that safe:

1

Upload

Your domain owner drops in the documents you already maintain: spreadsheets, transcripts, SOPs, standards. No reformatting, no special templates.

2

Parse

File contents are extracted server side. Nothing executes client side, and nothing leaves the deployment except the model call.

3

Structure and validate

A proprietary structuring and validation layer converts raw content into the helper's knowledge format, and rejects anything malformed before it can ship. Bad input cannot reach production. Stage 3 internals are licensed IP.

4

Human preview

Your owner reviews the structured result on screen before anything goes live. Nothing publishes without a human decision.

5

Versioned publish

One click writes a new version to the knowledge store. The helper serves the new content within about 30 seconds. The prior version remains recoverable.

Governance property. The model can only know what passed validation and human review. Its knowledge has a change log, an owner, and a rollback path.


Section 09

Per-module data classification. What touches what

Knowledge hubs

Internal methodology: rubrics, procedures, customer requirement documents, training transcripts. No customer records, no personal data. Hosted during evaluation, then inside your tenant in production.

Workflow tools

User inputs such as quantities, options, and assumptions, plus vendor and internal-contact business data. Anything your policy classes as confidential stays inside. Evaluation uses sample data only. Real data runs inside your tenant, never on external hosting.

Integration

Writes confirmed planning records into your ERP or equivalent. Final phase, inside your tenant, coordinated with your IT. Never from the evaluation deployment.

The never list, across every module

No confidential records leave your environment

Stated in the engagement agreement and enforced by architecture. Confidential inputs only ever run inside your tenant.

Model API keys never reach the browser

Server-side environment variables only. No client bundle contains them.

The model provider does not train on your traffic

Commercial API terms. Contractual, not a toggle. Enterprise zero-retention available where required.

No writes without a human confirm

The system-of-record write happens only after a person confirms the form, under a least-privilege service account, and every write is logged.

Conversations are not stored

No chat database. The audit log records actions and writes, not the content of questions.


Section 10

Security model, layer by layer

Layer Evaluation deployment Production in your tenant
Transport TLS 1.2 or better on every hop: browser to server, server to model API, server to knowledge store. Unchanged.
Access Password gate issuing an httpOnly session cookie, scoped to your evaluation group. Your SSO (Entra ID, Okta, Google). The app is built to swap gates without rearchitecting.
Secrets All keys in server-side environment variables on the host. Nothing in source, nothing client side. Same pattern inside your tenant, under your key management.
Source code Private repository. Copyright notices in every file. Access individually granted. Licensed artifacts transfer per the engagement agreement.
Storage Versioned knowledge store holds curated guidance only. No personal data, no data from your systems. Unchanged in kind. Data residency configurable to your policy.
Integrations None. The evaluation deployment has no credentials to any of your systems. Nothing to leak, nothing to pivot from. Final phase only: form to system of record via a least-privilege service account, write after confirm, every write logged. Scoped with your IT.
Revocation Kill switch. Hosted access, so all access can be revoked same-day from the deployment. Control transfers to your IT with the production handoff.

Honesty as the security posture. The evaluation gate is deliberately evaluation-scope. It is appropriate for a gated demo of non-sensitive content, and it is named as such rather than dressed up.

Your security reviewer should find every claim here exactly as strong as it is, and no stronger. A pre-answered security questionnaire is available on request.


Section 11

Not a product you configure. A system built for you

Custom

The corpus

Your standards, procedures, transcripts, and worked examples. Curated with your domain expert, not scraped.

Custom

The rules

Scope boundaries, refusal behavior, citation style, and escalation owner, tuned to how your reviewers actually judge answers.

Custom

The gate and the tenant

Your SSO, your cloud or Copilot environment, your data-residency policy for production.

Custom

The voice and the brand

Your name on it. Suggested prompts, tone, and UI written for your users' vocabulary.

Custom

The test pass

An adversarial question set built from what your people actually ask, run before every demo and every knowledge update.

Proven

The machine underneath

The two-layer grounding architecture, the knowledge pipeline, the security model, and the documented handoff path. Proven once, reused for you.

This is why a custom build lands in weeks, not quarters: the machine is done. The engagement is the knowledge work.


Section 12

Hosted to prove it, handed over to run it

Ph 1

First knowledge hub

Your highest-pain rubric or standard. Proves the pattern on your content. In the reference deployment the second hub then shipped in days on the same platform.

~3 wks
Ph 2

Second hub

Same machine, new corpus. Typically customer or regulatory requirements. First fully paid phase.

2 to 3 wks
Ph 3

Unified calculator

Your spreadsheets become one chat-and-form hybrid. Consumes the cross-link tags the first hub already carries.

6 to 10 wks
Ph 4

Supporting hubs and tools

Remaining hubs, vendor directory, contacts routing. Each on the same platform.

6 to 8 wks
Ph 5

System-of-record integration

The only outbound write in the system. Inside your tenant, with your IT.

3 to 4 wks
Ph 6

Knowledge transfer and IT handoff

Walkthroughs, documentation, and deployment support until your IT runs it alone.

2 wks

The production path

Each phase stands alone. Every phase is a working deliverable on the shared platform. Stop after any of them and keep everything built so far. That is the architectural reason the platform layer comes first, not a sales line.


Section 13

What we tell you before you ask

What is actually built

In the reference deployment, two of nine modules are live, each with chat, structured tools, and a browsable reference. The rest exist as clickable previews: amber-bannered, watermarked, sample data only. Previews are deliberately impossible to mistake for working software. Every terminal action is intercepted with a "just a preview" notice. They are designed and priced per engagement, not built.

Access during evaluation

Shared password, evaluation group only. No integrations, no confidential data. Production swaps in your SSO. Calculators and system-of-record writes run only inside your tenant.

Latency, measured

8 to 25 seconds is typical for a cited answer in the reference deployment. Published numbers come from live production tests, not lab estimates. Yours get measured the same way.

What this page is, and is not

This describes behavior, not blueprints. The prompt architecture, the knowledge schema, the cross-link tag design, the structuring and validation internals, and the curation methodology are licensed intellectual property. They transfer with a formal engagement, not with a page and not with a demo login.

That is not evasion. It is the same reason this can be shown to any audience safely. What you can evaluate is what it does, measured and in production. What you license is how.

Built to be auditable, not magical.


Section 14

Where this pattern fits. It is a shape, not a sector

The qualifier is not your industry. It is a shape with four parts, and if all four are present the pattern fits you regardless of what you make or sell.

1

There is an authoritative written thing

A standard, spec, rubric, policy, code, or manual. It exists and it is correct. Without this there is nothing to ground against, and you have a documentation project first.

2

There is a gap between the document and the decision

Reading the standard does not tell someone what to do at 2:15 on a Thursday, three fields into a form, with a deadline.

3

There is a bottleneck expert

One or two people who can close that gap, and everyone already knows their name.

4

Getting it wrong costs something real

A failed audit, a violation, a rework, a claim, an injury, a lost contract. Without this nobody will fund it, and they are right not to.

Manufacturing & supply chain

Reference deployment

"What evidence does the auditor want for criterion 3.2?"

Audit and supplier standards. Self-assessment with a printable gap report.

Biotech & life sciences

Gowning & deviations

"What does gowning qualification require before I can enter Suite B?"

GMP SOPs and cleanroom procedures. Deviation checklist escalates to QA in Teams.

Pharma & medical devices

Adverse event reporting

"Does this complaint need to be reported as an adverse event?"

Quality-system procedures. Decision-tree wizard that refuses when the case is genuinely ambiguous.

Healthcare providers

Policy at the bedside

"This patient meets two criteria that point different directions. Which policy governs?"

Infection control, medication, documentation and consent policies. Cited answers, paged escalation when the policy is silent.

Financial services & insurance

Pre-bind and appetite

"Can we bind this policy without a flood certificate?"

Underwriting guidelines and compliance manuals. Pre-bind checklist, unknowns routed to the named owner.

Energy & field service

Isolation and permits

"What are the isolation steps before I open this panel?"

Safety and lockout-tagout SOPs. Inspection checklist ending in your own STOP card on any failure.

Aerospace & defense

Stacked requirements

"Which requirements flow down to this supplier, and which do not?"

Your quality manual plus each customer's supplier requirements. Two corpora, answered where they interact. Controlled material stays in your tenant.

Food, beverage & agriculture

Deviation while the line runs

"A control point ran out of spec for eleven minutes. Deviation, hold, or recall?"

HACCP plans, preventive controls, allergen program. Digitized deviation record your audit scheme already requires.

Construction & engineering

Spec against code

"The spec says one thing and the local amendment says another. Which governs?"

Project specs, code amendments, and your own RFI history, which is a documented interpretation nobody can find twice.

Professional services

Policy, not judgment

"Does this engagement trigger an independence issue?"

Methodology, independence policy, engagement acceptance. Frees the judgment work, which is the part clients pay for.

Public sector & higher ed

Grant allowability

"Is this cost allowable on this award?"

Procurement policy plus the terms of each specific award. Different terms per award is exactly what a versioned multi-corpus store is for.

Logistics & transportation

Which rule applies

"Does this shipment classify as hazardous, and what does that change?"

Carrier requirements, routing guides, dangerous goods procedures. A scope wizard rather than chat: decisive, cited, fast.

And where it does not fit. When there is no authoritative document, there is nothing to ground against. When the question is pure judgment, a grounded system should refuse, so you paid for a refusal. When the question is a record lookup, your system of record already answers it. And when nobody will own the corpus, it goes stale within a year and people stop trusting it.

That last one is the only real prerequisite, and it costs nothing to check before you start.


If you are taking this to a steering committee

The six things they will ask

Want the fill-in-the-blank version of the one-page internal memo? Email me and I will send it, along with the full architecture deck including the limits page.

Bring me the standard that burns the most hours

About three weeks later you have a working knowledge hub on your content: grounded chat with citations, a structured self-assessment, and a browsable reference. You keep the system, the corpus, the tests, and the code either way.

Get in Touch Read the Case Study Read the White Paper