Client implementation · Production delivery

Delivering a Bounded AI System: Discovery, Deployment and Recovery

Visitors to a services business needed reliable answers and the right artist and booking path. The owners needed manageable enquiries and a system they could understand and maintain. I led discovery and delivery of a bespoke TypeScript implementation with two bounded model roles, application-enforced controls, evaluation tooling, and live client training.

My responsibility
Discovery, architecture, implementation direction and review, release, testing, and ongoing support
Delivered
Production website/AI integration, evaluation tooling, operating controls, bespoke guides, and live client training
Technology
TypeScript, OpenAI API, structured outputs, Wix hosting, and Node.js evaluation/operations tooling

Start with the business workflow and customer-data map

The engagement covered the website, an external scheduler, customer enquiries, and CRM follow-up design. Visitors needed approved service information and a reliable next step. The owners needed an experience they could explain, update, and recover when something failed.

I mapped where information originated, where customer data would travel, and which system could confirm an outcome. That included consent and data-minimization requirements, enquiry attribution, and retention decisions. A booking-link click, a saved enquiry, and a confirmed appointment needed separate meanings in the design.

My advertising-platform work at Yahoo, Criteo, and Meta informed this discovery. Integration work had already taught me to ask who owns a signal, what an event actually proves, and which downstream dependency can invalidate a reassuring dashboard.

Give models a bounded decision contract

The delivered system uses two model roles with distinct instructions and structured outputs. They propose approved fact IDs and actions. Requests take zero, one, or two model calls, depending on the path. Known direct-answer paths can stay in application code; the second role is invoked when its specialized decision is needed.

The TypeScript controller validates the proposals, checks allowed values, resolves the relevant approved information, composes the customer-facing response, and controls booking/contact destinations. Business information is available in English, Spanish, and Vietnamese. Selected public prices and provider listings are retrieved from the scheduler when needed.

  1. Request and context: validate the input and determine the supported path.
  2. Model proposal, when needed: return structured intent, approved fact IDs, and permitted actions.
  3. Application checks: validate the schema and allowed values; resolve approved copy and selected live data.
  4. Controlled response: compose the wording and provide a permitted destination or human-escalation route.

This division made the decision boundary explicit. The application remained responsible for the published wording and the action it offered. I used AI coding tools during implementation and documentation, while retaining responsibility for scope, design decisions, review, release, and verification.

A returning-client price question, end to end

Consider a visitor asking, “How much is a returning appointment with [artist]?” This is illustrative wording for a verified implementation path. The artist is anonymized; no customer message or current price is reproduced.

  1. Recognize the supported request. The controller identifies an explicit returning-client price question and a supported artist, after checking the existing escalation boundaries. This path uses zero model calls.
  2. Look up the source. The pricing adapter retrieves the selected public scheduler listing and checks its supported structure and required conditions.
  3. Compose a validated answer. Application code presents the available follow-up and annual-maintenance prices with their eligibility conditions, and distinguishes the deposit from the full service price.
  4. Offer the correct next step. The response links to the returning-client booking route with the verified artist selection. Opening that route establishes navigation; a confirmed appointment requires a separate scheduler record.

Missing, stale, ambiguous, or unsupported listings take the native-listing or human-contact fallback. The implementation supports a defined listing shape; broader price coverage would require its own parser support and verification.

Define what each event can establish

The scheduler remains a separate system. The website can observe a visitor opening a booking route, while the native enquiry path can confirm that a submitted message was saved. Confirming an appointment or revenue requires a separate downstream record and an appropriate integration.

SignalEvidence availableInterpretation
Booking/contact link activityBrowser-reported actionReported intent or navigation; downstream completion remains unknown.
Native enquiry submissionServer-confirmed saveThe enquiry was recorded; staff response and a booking remain separate outcomes.
Confirmed appointment or revenueRequires a verified downstream recordA booking/revenue join was outside the implemented measurement scope.

CRM assessment and follow-up workflow design were part of the consulting work. Activating automated CRM follow-up remained separate scope. Keeping those boundaries visible helped prevent the interface, analytics, and operating instructions from promising different things.

Separate model evaluation from software regression coverage

I built Node.js tooling around authored synthetic scenarios to inspect model decisions and end-to-end behavior. The final October 9 development run passed 46 of 46 scenarios covering pricing, service/provider routing, and escalation boundaries across English, Spanish, and Vietnamese. Thirty-four scenarios exercised real model decisions and twelve tested deterministic boundaries; the run made sixty model API calls across the two roles.

Separately, the October 26 verification passed 284 software regression checks: 276 concierge-project offline tests and eight production-adapter routing/proxy tests. These checked implementation behavior such as validation, response composition, and control paths. They are separate from the model-decision evaluation.

The 46-scenario suite was used during development and passed after fixes, including a routing error between similarly named services. Its result applies to the evaluated revision. Later client feedback exposed additional paraphrase gaps. Held-out performance, production accuracy, broad safety, and commercial impact require separate evidence.

Investigate the path that failed in production

After launch, an owner-reported failure exposed a gap in the operational checks. Reads and configuration checks could succeed while required quota-admission writes were rejected because expired records had accumulated in the hosted data store.

I diagnosed the failing write path and coordinated a backed-up, expiry-checked recovery that removed 523 expired quota records. All 301 retained quota records and the control configuration were verified unchanged; active spending reservations were preserved. These were technical quota records. A separate parser-coverage defect affected a returning-client price listing. I corrected it and verified the reported wording through the affected customer path.

Protected scheduled retention was then deployed through the hosting platform. Controlled deletion and a later natural scheduled execution were verified separately. The incident sharpened the monitoring requirement: the health of a read endpoint and the ability to admit a real request need separate checks. Independent missed-run/failure alerts remain follow-up work.

Make operation understandable to the client

The implementation included input validation, server-side credentials, durable rate and spending controls, a shared pause mechanism, human escalation, and private diagnostic traces. The operating documentation explained architecture, limitations, troubleshooting, ownership, and maintenance.

I also delivered bespoke illustrated website/AI walkthroughs, account-setup and funding guidance, and live one-on-one explanation and setup assistance. That work translated implementation decisions into tasks the business owners could understand and use.

Training delivery, demonstrated unaided proficiency, and accepted ongoing operational ownership are separate milestones. Continuing ownership/access acceptance remained open at the documented checkpoint, with me still providing support.

What this implementation demonstrates

The website and two-role AI system reached production on the business domain in October 2026, using the business's model account. The engagement produced a working customer-facing implementation, evaluation and operations tooling, production diagnosis and recovery, customer-data design, and delivered client enablement.

Customer-traffic volume, conversion lift, time savings, and revenue impact have not been measured for this engagement. The demonstrated work is the implementation and its verified operating steps. The next evaluation questions include held-out routing behavior, source-freshness failures, and the independent operation of maintenance and escalation workflows.

This is the combination I bring to applied AI and solutions engineering: understanding the customer's systems, designing explicit integration contracts, getting an implementation into production, investigating its failures, and helping people operate it.