From reading docs and filing tickets, to getting the right answer in seconds. An AI chat assistant built for developers integrating with the Azuga Fleet API.
Developers integrating with the Azuga Fleet API faced a frustrating loop:
- Read through lengthy API documentation
- Still unsure which endpoint fits their use case
- Write to support and wait for a response
- Repeat
Support tickets were piling up. Integration timelines were stretching. The documentation existed — but it wasn't findable in the moment developers needed it.
The question: What if developers could just ask?
A conversational AI assistant embedded directly on the Azuga Developer Portal. Developers type their question in plain English and get back:
- The exact API endpoint for their use case
- Request and response structure for that endpoint
- Guidance on migrating from older API versions
- Instant answers — no ticket, no wait
A sample
Built on Chatbase using a RAG (Retrieval-Augmented Generation) architecture:
- Entire developer portal webpages and API docs uploaded as the knowledge base
- Chatbase indexes and embeds the content
- Developer asks a question in the chat widget
- Model retrieves the most relevant documentation chunks and generates a precise answer
- Response is scoped strictly to the uploaded knowledge — no web search, no hallucination
This is where product thinking drove the build. Every guardrail was a deliberate decision:
| Guardrail | Why |
|---|---|
| No web search — only uploaded docs | Prevent hallucinated endpoints that don't exist |
| Return only specific endpoints | Avoid generic, unhelpful answers |
| Include request + response examples | Developers need the full picture, not just endpoint names |
| Custom answer for auth key generation | Common question with a specific, non-obvious answer |
| Redirect older API version questions to latest | Drive migration, reduce support for deprecated endpoints |
| No hallucination instruction in system prompt | Explicit guardrail — if unknown, say so |
North Star Metric: Reduction in API-related support tickets
Supporting metrics:
- Chatbase adoption rate (active users / portal visitors)
- Helpfulness rating per response (thumbs up/down)
- Answer accuracy rate
- Support ticket volume (before vs. after)
- Query resolution time
| Component | Tool | Cost |
|---|---|---|
| AI Chat Engine | Chatbase | Free tier |
| Knowledge Base | RAG via uploaded webpage links | Free |
| Embedding | Chatbase built-in | Free |
| Integration | Embedded widget on developer portal | Free |
- Reduced integration query resolution from 1 week → under 1 hour
- Measurable drop in API-related support tickets post-launch
- Increased developer portal engagement 60×
- Feedback loop — use thumbs down ratings to identify documentation gaps and improve them
- Intent classification — detect when a question is outside scope and route to human support automatically
- Multi-language support — serve non-English speaking partner developers
- Analytics dashboard — track most asked questions to inform documentation prioritisation
- Upgrade to Claude/GPT — move beyond Chatbase for more nuanced, multi-turn conversations
This is a PM case study and system design. Built as an internal product at Bridgestone/Azuga. Details shared for portfolio purposes.