DoggieAI

Doggie Education’s digital support service: practical help between lessons, with clear limits and a direct path to an educator or veterinarian.

Comment

Inspiration

DoggieAI was already the name of this service before Build Week. It comes from a practical need I see through Doggie Education: owners do not only have questions during a scheduled lesson. Doubts appear on a noisy walk, after a difficult interaction, or when a dog suddenly behaves differently.

I wanted to offer continuity in those moments without pretending that a chatbot can replace an educator or a veterinarian. An early prototype already existed. For OpenAI Build Week I used Codex with GPT-5.6 to examine that real codebase, find weak points and turn it into a safer, more complete product.

What it does

DoggieAI is the digital support service of Doggie Education. It asks progressive questions about the dog, the situation and what has already been tried, so the conversation can continue instead of starting from zero every time.

There are two deliberately separate paths. The public chat answers questions about services, prices and bookings. Authenticated subscribers can receive personalized educational guidance, with conversation history, message quotas, account management, extra-message purchases and Stripe billing.

The service also knows when it should stop. It directs emergencies and veterinary matters to the appropriate professional, and it can hand a case to the human team through WhatsApp or an encrypted callback request. Users are clearly told that the AI can make mistakes.

Behind the chat there is a curated knowledge base, verified service links and an administration area for conversations, documents, plans, subscriptions, leads, consent records, discount codes and AI cost monitoring.

How I built it

I chose a lightweight stack because the product has to run on ordinary shared hosting, not only in a hackathon environment. The application is server-rendered PHP with MySQL/MariaDB, HTML, CSS and JavaScript. It runs on Hostinger and includes internal REST-style endpoints for chat, subscriptions, account activity, feedback and human handoff.

The production AI layer currently uses Gemini-family generative models and embeddings with retrieval from the curated knowledge base. Stripe Checkout, signed webhooks, idempotent event handling and Customer Portal manage subscriptions.

During Build Week, Codex with GPT-5.6 became my engineering partner. I used it to map the architecture, review and refactor roughly 130 PHP files, harden authentication and tokens, make quotas and payment events concurrency-safe, repair migrations and the updater, improve retrieval boundaries, redesign the mobile experience and validate PHP 7.4 and JavaScript syntax.

I set the product requirements and made the final decisions about behavior, safety, privacy and user experience. Codex helped me test those decisions against the whole codebase and implement them consistently.

Challenges I ran into

The hardest issue was separating free service information from personalized dog-training guidance. Prompt instructions alone were not reliable enough. I therefore added application-level classification and limited which knowledge documents the free mode is allowed to retrieve.

A second issue was factual accuracy. Semantic search cannot guarantee the exact booking link or commercial detail on every request. I created a server-side registry of 52 verified official URLs and deterministic routing for 23 specific services; any service link outside the registry is rejected.

Other less visible problems mattered just as much: simultaneous requests, duplicated or delayed Stripe webhooks, expired events, model-provider overload, PHP 7.4 constraints, mobile overlays, cookie controls and recovery from failed database updates.

Accomplishments that I am proud of

DoggieAI is now an end-to-end service rather than a chat window. The public assistant, subscriber consultations, persistent profiles, payments, account management, human handoff, back office and privacy records work as parts of the same product.

What matters most to me is that important rules are enforced by code instead of being left to the model: access rights, payment amounts, discounts, verified URLs, service boundaries and dangerous cases.

What I learned

My main lesson is that a trustworthy AI product is mostly built around the model. Retrieval quality, deterministic safeguards, billing integrity, privacy, monitoring and a clear route to a human professional are at least as important as the generated response.

Codex was especially useful when I treated it as a collaborator for inspection and testing, not as a one-click project generator. The strongest results came from repeated passes: reproduce a problem, challenge the proposed fix, test it against existing behavior and review the final change.

What's next

Next I will build a structured evaluation set for common dog-behavior scenarios, improve model routing and cost monitoring, complete the Stripe test matrix, move the hosting runtime to a supported PHP version and connect DoggieAI with DoggieHub through permissioned, auditable APIs.

Built With

Share this project:

Updates