Conversation services, retrieval, streaming APIs and the platform that keeps them observable in production. Ten years of shipping Python.
I like owning a service end to end: the architecture, the asynchronous Python, the data model, the CI pipeline, the Kubernetes deployment and the dashboards that tell you it's healthy.
Ten years in production software across telecoms test automation at Ericsson, unified communications at Comcast, applied computer vision, and most recently LLM conversation and retrieval systems serving live customers. I hold an M.Sc. in Cognitive Science and AI from Seoul National University and have published at CVPR Workshops and CogSci.
I care about typed, tested code and about being the person who reviews it.
BEN builds LLM-powered conversational assistants with speech for customers. My work sits in the AI backend: the services that turn a user's utterance into a spoken, grounded, context-aware reply, and the platform that runs them.
Architected the asynchronous Python library that all of the company's dialog managers build on: text generation, embeddings, emotion classification and conversation state.
Decoupled speech synthesis from dialog logic across six customer assistants, so multi-sentence responses are synthesised in parallel. That cut response latency by roughly 40% and simplified deployment.
Designed and shipped a streaming chat API over WebSocket and Server-Sent Events with typed event frames, automatic language switching, per-voice text-to-speech routing and multilingual fallbacks.
Brought the service to a typed, linted codebase with an 80% coverage gate in CI.
Built the retrieval service on PostgreSQL and pgvector: per-tenant databases, pluggable dense and sparse embeddings, background re-embedding and parallel query decomposition.
Co-designed metadata-aware reranking and table and image retrieval from scientific papers, with RAGAS-based evaluation to measure changes.
Led a coordinated breaking change that replaced direct database access in five microservices with a single typed data-service client.
That made schema ownership explicit and unblocked the move to a multi-tenant deployment.
Added end-to-end trace IDs, Prometheus metrics and structured logging with Loki and Grafana across the AI stack.
Own the Kubernetes deployments and releases for the services I work on, including an event-driven Celery task service for conversation, summary, retrieval and analytics persistence.
Co-designed a memory system that builds long-lived user profiles from conversations through per-turn LLM extraction over Redis-backed sessions, so assistants remember what matters across sessions.
Built Streamlit chat, speech and retrieval testers that engineers and domain experts use to try changes before they ship.
Backend developer on Comcast's Unified Communications platform. RESTful APIs with Flask and Marshmallow, asynchronous task processing with Celery, RabbitMQ and Redis.
Built a sharded, distributed database in Node.js on a Proof-of-Stake protocol: highly partitioned, always available and eventually consistent, tuned for transactions per second.
Published on hierarchical colour learning in CNNs, spoken command classification and motif discovery. Edited lab members' publications on the Visual Turing Test.
Thesis: Investigating Hierarchical Color Learning in Convolutional Neural Networks, supervised by Prof. Byoung-Tak Zhang. Transcript
Exchange year at the University of Hong Kong. Transcript