CASE STUDY 03 — REAL-TIME AI
AI Conversation Platform
As an Application Developer in MarConPra's Media Management Division, I built a real-time AI conversation product: it captures live speech, transcribes it, reads the tone, and streams back AI-suggested responses in under a third of a second — all delivered to a React Native console over WebSockets.
OVERVIEW
Turning live speech into intelligent, real-time responses.
The product listens to a live conversation, transcribes it as it happens, classifies the speaker's tone, and generates helpful AI responses that appear on screen almost instantly. My focus was architecting the streaming backend and the pipeline that ties speech, transcription and language models together, and delivering it through a cross-platform mobile console.
Because it was used live in front of stakeholders and prospective clients, everything had to feel immediate and stay reliable — no awkward pauses, dropped connections or stale transcripts.
HOW IT WORKS
The real-time pipeline.
Every stage runs asynchronously so audio, transcript and AI output move through the system without blocking one another — which is what keeps end-to-end latency under ~300ms.
WHAT I BUILT
The pieces behind the product.
Streaming backend
FastAPI and WebSocket services to stream speech, AI responses and real-time updates with bi-directional, event-driven communication.
Speech pipeline
Live audio capture and streaming transcription via AssemblyAI, with synchronized, continuously-updating transcripts.
LLM integration
Transcripts run through LLMs for tone classification and response generation, with prompt/context engineering and multi-turn handling.
React Native console
A cross-platform React Native + Expo console for driving conversations, showing live transcripts, tones and AI suggestions.
Session & connection handling
Connection management, auto-reconnection and session handling so a live demo keeps flowing even through network hiccups.
Latency optimization
End-to-end pipeline tuning across capture, transcription and LLM inference — cutting AI processing time by ~30%.
ENGINEERING HIGHLIGHTS
What made it work in real time.
- Architected FastAPI + WebSocket services to stream speech, AI responses and updates across 30+ live product demos.
- Built an AI-driven React Native console with Expo, integrating AssemblyAI transcription and LLM tone classification.
- Implemented 15 multi-role conversation flows with speaker context, role switching and conversation tracking.
- Reduced AI processing time by ~30% through async execution and end-to-end pipeline optimization.
- Kept end-to-end response latency under ~300ms for a natural, live feel.
- Handled production realities — WebSocket disconnects, AI-response issues and mobile bugs — with root-cause fixes before demos.
ROLE
Application Developer, Media Management Division.
I worked across backend, real-time systems, AI integration and the mobile console, collaborating with AI engineers, designers and the product team to take features from idea to a working, demo-ready product — and to keep them running smoothly when it mattered most, live in front of an audience.
TECH STACK
What it's built with.
Python, FastAPI, Fastify — REST APIs and async services
WebSockets, event-driven streaming, connection management
AssemblyAI streaming transcription, live audio processing
LLMs for tone classification and response generation; prompt & context engineering
React Native, Expo
Async programming, performance optimization, production debugging, Git
A NOTE
This was internal product work at MarConPra. Some details are proprietary and intentionally left out here — I'm happy to walk through the architecture and my specific contributions in a conversation.