Technical record
Tim: useful website answers within a clear boundary — technical record
Architecture, tests, review findings and historical scope supporting the main case study.
Written with assistance from Dottie, Mike’s AI dot assistant, under Mike’s direction and review.
Tim is a small character that helps visitors explore Iron Walker Studios. His answers are written and reviewed ahead of time, matched to questions in the browser and linked to the website pages that support them. The release candidate uses this curated guide.
Set a useful boundary
I wanted the helper to answer questions about the website, preserve the conversation when reopened and remain unobtrusive around the artwork and playable game. I also wanted a firm budget boundary and a clear distinction between public content and private development material.
I set the behavior, wording, visual direction and acceptance requirements, then used AI-assisted coding to implement and revise the guide. I retained the decisions about what it should say, what evidence belongs on the public site and whether a result is ready to release.
Explore the API approach before choosing the delivery
I directed a server-side prototype with a public-page source boundary, persistent visitor quotas and atomic cost reservations. The endpoint, provider boundary and ledger contracts were implemented. Deterministic preview responses and simulated provider results exercised concurrency, replay, timeouts, uncertain outcomes and renewal. Local database checks also tested persistence and restart behavior. These were local prototype checks, not requests to a live model or a deployed service.
The budget target was $5 per month. Provider alerts and browser counters would not enforce that limit. The prototype used conservative reservations and failure handling, but production enforcement still needs a proven durable storage deployment, current pricing and token bounds, verified provider-usage settlement and review of unknown outcomes.
A live version would also need its own public-content release review, server-only credential and billing configuration, live grounding and refusal evaluations, leakage tests, privacy review and operational monitoring. Its real answer quality and production budget behavior remain acceptance work. The live AI service remains deferred and disabled.
I deliberately reduced this release to prepared browser-side responses: useful site answers, approved source links, retained conversation and the character interaction. That removes model calls, quota enforcement and database operation from the current delivery. Tim adds no paid AI service; ordinary website hosting and delivery retain their own costs. The implemented API prototype stays separate for a later decision.
Keep answers attached to public sources
Curated topics contain short answers, sample questions and matching terms. Each answer points to approved public website routes. The browser receives the answers and links, while supporting review quotes and source fingerprints stay out of its guide data.
A public-page text change requires reviewing the affected answer and its evidence before recording a new fingerprint. Shared limits keep source counts, URL lengths and titles consistent between content validation, live replies and reconstructed history. Tests cover the boundary cases so a supported content update cannot silently erase its citations.
Matching is deliberately limited. Clear topic buttons give a dependable route to an answer; ambiguous questions offer choices, and unsupported questions return useful site links. Tim does not understand arbitrary meaning, browse the web, execute tools or invent a response to fill a gap.
Make the character useful without interrupting the page
On Home, Tim appears after the header fully collapses, with one magical swirling poof and an introduction once per tab session. On other pages he appears without that introduction. Opening him reveals the conversation; closing returns him to the compact character.
Keyboard activation, focus return and Escape support the same flow as clicking or tapping. Pointer movement and a restrained scroll reaction give the character some life. Reduced-motion preferences suppress distracting animation, and the guide hides while the game player is loaded.
Messages render as text with validated source links. Questions, answers and retained history have size limits. No generated HTML from a response is inserted into the page.
Preserve conversation without creating a server record
Tim keeps up to 60 messages from the latest 30 days in this browser. Expired records are removed when the guide next runs; storage does not delete them automatically while the site is unused. Closing Tim keeps it available when reopened or reloaded at the same origin. Clear conversation is separate from appearance preferences and game saves.
Where browser coordination is available, ordered Clear operations protect newer messages from sibling tabs. If safe coordination is unavailable, Tim explains that it is using this tab only and leaves saved or sibling history untouched. Private browsing, cleared site data and a different site address can remove access to earlier conversations.
This mode has no visitor identifier, new cookie, IP safeguard or server-side chat history. The privacy page describes the actual storage behavior.
What this work demonstrates
The result is a focused website guide with inspectable answers and an explicit operating model. Content checks, artifact checks and UI regressions exercise different boundaries; browser interaction and network review remain separate release checks.
The tradeoff is concrete: prepared answers offer predictable content and deployment at the cost of broad conversational flexibility. A later live AI version would go through separate cost, grounding, privacy and operational review.
Value to an employer
I treated the helper as a product and operating decision: what visitors need, which sources it may use and what the budget can sustain. I directed a curated browser guide with source links and local history, alongside a separate API prototype that exercised quota and accounting contracts with simulated provider responses.
For an Applied AI team, this offers examples of cost-aware tool selection, explicit failure handling and deliberate scope reduction. The useful outcome is a guide that can be inspected and reviewed now, with a documented path for evaluating a live model later. Current prepared answers, a tested local prototype and production AI operations have separate acceptance requirements.
Evidence boundaries for the guide and prototype
The guide’s prepared answers, source validation, local history and browser behavior have their own checks. Simulated provider responses in the separate prototype exercise contracts without establishing real model quality, production cost enforcement or deployment.
The recorded defect traces cover the specific findings described here. Other regression boundaries can be extended with a documented finding, root cause, correction and verification as issues are discovered. The guide is the demonstrated delivery; the live API remains deferred.