Case Study
AI Instructional Design Platform
PotSolutions designed and built a multi-tenant SaaS that guides instructional designers through the ten-stage Criterion-Referenced Instruction (CRI) workflow. At every stage Claude analyses or drafts, a rubric scores the result, and a person accepts it — so the methodology stays in charge, not the model.
The Challenge
A Proven Methodology, Not Another Chatbot
The client is a training-methodology company whose Criterion-Referenced Instruction approach has been used to design professional training for decades. They wanted to turn that methodology into a licensable AI product.
Generic AI tools produce training content that reads well but ignores the method: vague objectives, practice that doesn’t match the skill, and modules padded with whatever was in the source documents. Enterprise customers also needed to know their company documents stayed private, that one customer could never see another’s data, and that their own identity provider controlled sign-in.
The design principle agreed with the client was simple and strict: define the objective, the practice, and the content the practice needs before any company documents are uploaded. That order is the primary filter against fluff. The AI proposes; a person decides at every step.
PotSolutions owned the engagement end-to-end — product architecture, the AI workflow and prompts, the full-stack application, the Azure estate, CI/CD, security, and the beta-readiness plan.
The Workflow
Ten Guided Stages, One Person Accepting Each
Each training module moves through the CRI sequence. Every stage after setup has one AI tool, and later stages read only what earlier stages accepted.
| # | Stage | What the AI produces |
|---|---|---|
| 1 | Module setup | Context, audience and constraints for the module (form or chat) |
| 2 | Objective | Evaluates and rewrites objectives against a rubric, up to three improvement passes |
| 3 | Skill check | Drafts the check that proves the objective was met, scored against a code rubric |
| 4 | Practice & feedback | Practice exercises and relevant feedback, scored per item |
| 5 | WIIFM | “What’s in it for me” — why the learner should care |
| 6 | Derive content topics | Only the content the practice actually requires |
| 7 | Job aid analysis | Module only, hybrid, or job aid only — a job aid can replace a whole module |
| 8 | Sources & content fence | Scoped document upload, then what the module must and must not include |
| 9 | Draft content | Authoring instructions grounded in the fenced source facts |
| 10 | Skills audit | Completeness audit of every accepted artifact against CRI criteria |
What We Built
An AI Product with Guardrails Built In
The workflow is enforced by the server, the AI is grounded in the customer’s own documents, and every piece of text the platform didn’t write is treated as data, not instructions.
Guided workspace and stage machine
- Server-authoritative stage machine — pure accept and reopen gates that return every reason a stage is refused, in a sentence the designer can act on. The UI never guesses gate state.
- Only a person accepts — the chat assistant can propose accepting a stage, but acceptance is always a human click, recorded with who accepted which artifact and when.
- Drift detection — every artifact records which upstream artifacts and sources it read, so designers are told when an earlier stage or a source document has changed since.
- Live streaming — tool runs stream progress over Server-Sent Events and every event is persisted, so a run can be replayed exactly as it happened.
AI tools, chat and grounded retrieval
- Code and model working together — the objectives tool segments text, flags vague verbs and scores in code, using the model only for structure and rewriting; it iterates until every rubric criterion passes.
- Refuse rather than invent — a model reply that can’t be validated becomes a visible refusal instead of a plausible-looking default.
- Semantic Kernel chat orchestrator — conversational control of the workspace, with the functions the model can call scoped to the current stage.
- Runtime model selection and metering — platform admins choose the Claude model without a redeploy; every model call is written to a token ledger for quotas and usage reporting.
Document ingestion and prompt-injection defence
- Queue-driven ingestion — uploads land in Blob Storage and are processed by a worker that KEDA scales from zero on Service Bus queue depth, so bulk uploads never slow the interactive streams.
- Cost-aware extraction — local readers handle PDF, Word and PowerPoint text first; Claude vision OCR is the fallback for scans and photos only.
- Structured fact inventory — documents become typed facts with page and quote, indexed in Azure AI Search, so every drafted claim traces back to its source.
- Layered injection defence — hidden and invisible text is stripped, untrusted text is fenced in every prompt, Azure Prompt Shields screens the untrusted text of every model call, and suspect facts are kept for review but never retrieved.
Identity & Tenancy
Built for Enterprise Customers
Customers bring their own identity provider, departments get their own workspaces, and isolation fails closed.
Sign-in and authorization
- Four sign-in methods — Microsoft 365 (multi-tenant Entra ID), SAML 2.0, passkeys, and email codes; each customer tenant chooses which it allows.
- Corporations and tenants — one customer organisation can hold several workspaces, so tenancy lives in our own data and the identity token never selects a tenant.
- Role-based permissions — tenant admins, content managers, designers, subject-matter reviewers, viewers and billing managers, resolved from the database on every request.
- Accountable operator access — support staff enter a customer tenant only through time-boxed grants or short impersonation sessions, each written to the customer’s own audit log.
Isolation and auditing
- Entity Framework global query filters that fail closed, forced Cosmos DB partition keys, tenant-filtered search, and tenant-prefixed blob paths — another tenant’s resources simply return 404.
- Every mutation appended to a per-tenant, append-only audit log.
- Short-lived session tokens signed inside Key Vault, with hashed, rotating refresh tokens and reuse detection.
DevOps & Security
Terraform Is the Only Way In
Three Azure subscriptions — dev, stage and prod — built entirely by Terraform and deployed by Azure Pipelines. The CLI and portal are for reading only.
Infrastructure and pipelines
- Terraform-only change control — 16 modules covering network, data, AI, compute, identity, monitoring and Defender for Cloud; every plan reviewed and approved per environment.
- Promotion by pull request —
dev→stage→prod, with separate API, web, worker, infrastructure and Entra pipelines per environment. - Safe database releases — self-contained EF Core migration bundles run as a gated stage before the app deploys; application identities hold no DDL rights.
- Verified deploys — each release polls the health endpoint until its build checksum matches the commit.
- VNet-injected build agents — Managed DevOps Pools reach data planes over private endpoints, so no pipeline ever opens a firewall rule.
Network and secrets
- Private endpoints for SQL, Cosmos DB, Blob Storage, Key Vault and AI Search, with private DNS linked per environment.
- NAT Gateway for a fixed, allowlistable egress IP; internal-only worker.
- Managed identities everywhere — storage keys and SAS disabled, SQL Entra-only in stage and prod.
- Content-Security-Policy and security headers on the SPA; the API reachable only through Cloudflare, with WAF rules and rate limits added during beta hardening.
Beta readiness
Shipping an AI product to enterprise customers is more than code. PotSolutions ran an eight-week beta-readiness plan with the client: a persona test matrix covering every role across every sign-in method, a ten-module scored test plan including an adversarial run with injected instructions, a go/no-go checklist, and a legal handoff pack giving the client’s lawyer and insurer the facts on subprocessors, data location, retention and security.
Technology
What We Used
Application & AI
Platform & DevOps
The Result
A Methodology Turned into a Product
The client now has a licensable, multi-tenant AI product that applies their method to every module it helps design — with the security, tenancy and operational controls enterprise buyers ask for — heading into its first customer beta in December 2026.
Similar Challenge?
Building an AI Product, Not a Chatbot?
Whether you’re turning your own expertise into an AI SaaS, need guardrails around model output, or have to pass an enterprise security review — we design and build production AI platforms from architecture through CI/CD.