FPT Software · Interview AI · Speech & Conversation
Voice is a product workflow, not just an API.
My FPT work also included multilingual speech and interview-oriented AI services. The surrounding team system connected a browser-based conversation experience with interview backends and asynchronous AI jobs. The hard questions were how to keep turns, language, status, processing and evaluation coherent across those boundaries.
Why it is different
Interactive audio meets asynchronous reasoning.
Recording, transcription, interview prompting and conversation evaluation do not share the same latency or failure characteristics. A useful application must distinguish what the user said, what was transcribed, which turn is active, and whether evaluation or generation has completed.
The source-backed technology environment included web UI and Azure Speech integration, Java interview APIs, Python asynchronous OpenAI-based scenario jobs and operational deployment tooling. These are team-system components; I do not claim personal authorship of each one.
The application boundary
Four stages of a voice-interview session.
Conversation experience
Interactive interview UI, voice capture and speech-service integration
Interview services
Session context, question/scenario handling and authenticated application APIs
AI dialogue jobs
Asynchronous language-model calls, prompts, retry behavior and conversation evaluation
Result and review
Structured session outcomes and separation of processing state from what the user sees
What the experience demonstrates
Stateful conversational product engineering.
This work strengthened my experience with speech-enabled software, multi-turn dialogue, interview-specific prompt flows, structured evaluation, API contracts and asynchronous failure modes — capabilities relevant to copilots and voice agents beyond a single provider SDK.
No claimed custom ASR model training, independently measured word-error rate, speech latency, fleet/user scale or SLA is implied by this case study.
Related work