Building Production AI Systems

LLM Applications, RAG Pipelines & Document Intelligence — Hong Kong · Remote-Friendly

Most businesses I work with are not short of data. They are short of a system that turns that data into something actionable without requiring someone to manually read through it every day. That is the gap I build for.

Background

I am an independent AI engineer based in Hong Kong and mainland China, available for remote engagements globally. My background spans 7+ years in data and analytics roles, 2+ years building and deploying production AI systems, and a Master's in Business Analytics and AI from Vlerick Business School (Belgium).

The systems I build are end-to-end: data ingestion, LLM processing, retrieval architecture, delivery layer, and infrastructure. Not prototypes. Production systems with real users and real operational requirements.

What I Build

RAG Pipelines and Document Q&A

Organisations that handle large volumes of documents — policies, reports, contracts, filings — typically face the same bottleneck: the information exists but is not queryable. Finding a specific detail means opening files manually.

A production RAG pipeline changes this. Documents are ingested, chunked, embedded, and stored in a vector database. The result is a system where a natural language question returns a precise answer drawn from the full document archive — regardless of volume or age of the files.

Scheduled Data Pipelines and Automated Summaries

The same underlying architecture — scheduled ingestion, LLM processing, structured output — applies to monitoring and summarising information sources on a recurring basis. Configurable inputs, relevance filtering, and delivery via Telegram or email. The pipeline runs on a schedule; the output arrives without manual effort.

This is applicable wherever a team currently has someone manually reading sources and summarising them.

Memory-Enabled AI Systems

I also build AI applications where context persistence matters — systems that remember user history, learn from interactions, and retrieve relevant prior context at inference time.

HKSoka (hksoka.com/en) is a Claude-powered chat platform I designed and built end-to-end. Its core differentiator is a multi-layer RAG memory architecture:

This architecture is directly applicable to any business context where continuity across sessions matters — client relationship management, ongoing advisory workflows, or knowledge base assistants.

Infrastructure and Deployment

Production systems require more than good model calls. The infrastructure I build includes:

I have also built and maintained a production ML trading pipeline with automated signal generation, multi-seed validation, and live deployment on AWS — including full infrastructure ownership from data ingestion to model deployment.

Engagement Model

I work on a project basis. Deliverables are scoped clearly before commencement. I deliver working systems, not slide decks.

For businesses looking to automate document workflows, build internal knowledge bases, or deploy LLM-powered tools into existing operations — I am available for scoping calls.

Contact: smartai.hk+ai.consulting@proton.me
LinkedIn: linkedin.com/in/levi-innovation