AI & Automation for Hong Kong Startups

Most AI work fails between the demo and production. I help founders pick the use cases that pay, build them on a stack you can operate, and put the guardrails, evaluation, and cost controls in place before customers see them.

What it covers

  • LLM features in your product: retrieval over your own data, structured extraction, agents with tool access, and the evaluation harness that proves they work.
  • Internal automation: the support inbox, document intake, reconciliation, reporting, and the glue between SaaS tools, built so it runs without a person watching.
  • AI product strategy: where a model earns its cost, where a rule is cheaper, and what to build versus buy.
  • Model and vendor choice, routing, caching, and budgets, so the bill tracks usage rather than surprising you.

Built the way it will be run

I run an AI executive board inside LX Software: a set of role-specific agents with tools, budgets, approval gates, and audit trails that handle routine work in a real company every day. That system is also the base of ai-cxo. The patterns it taught, function calling with strict schemas, propose-before-act permissions, daily budgets, and plain logs, are what I bring to client work.

Delivery follows the same rules as any other production system: infrastructure as code, CI/CD, observability, and a rollback path. A prototype is welcome; a prototype in production without those is not.

How it runs

A prototype sprint is a fixed scope over two to four weeks: one use case, a working build against your data, an evaluation set, and a cost model. If it earns its keep, the retainer carries it into production with your team.

Guardrails are part of the scope from the first day: what the system may read, what it may write, what needs a human, and what is logged. For Hong Kong companies that includes PDPO data-handling questions and the customer questionnaires that follow.

FAQ

Which models and providers do you use?

Whatever fits the task and the data rules: hosted frontier models through a router for most work, smaller or self-hosted models where cost or residency demands it. The architecture keeps the provider swappable.

Can you automate an existing manual process?

Usually, and often without a model at all. The first step is mapping the process and finding the parts that are rules; the LLM handles the unstructured remainder, with a human on the exceptions.