Forward-deployed engineer
I get it out. Embedded with the customer, on their own data, one workflow live in weeks rather than a deck that impresses and dies. Currently leading engineering on an agent and knowledge platform in Hong Kong.
Specimen — hover to reveal
A pilot proves the technology works, everyone agrees it is impressive, and nothing reaches production. The way past it is the same shape every time, whether it is a tender comparison, an ad compliance gate, or an agent platform.
A clickable path through the real decision, seeded and backendless, is cheaper than the meeting it replaces. The output is not a report. It is software an executive clicks.
Where I did this
Li & Fung, Strategic Data Unit
Product owner of the document intelligence service and lead for AI development across the group. One extraction contract serving several business units, rather than a separate parser per team.
One service for the group, or every unit builds its own.
A demo wins the decision; only the customer's own data proves the value. Sample data hides the fields that do not match the documentation, the third of rows that are null, and the two contradictory answers that are each locally correct.
Where I did this
Sun Hung Kai Properties, Chairman's Office
Telco, customer profile, transaction, image and text data fused into one view, cut into hundreds of segments by lifestage, interest, behaviour and demographics, and handed to campaign teams as target lists.
A segment is only real if a campaign team can act on it.
Owning the pipeline and not just the model is the difference between a notebook and a system: sourcing, cleansing, features, the model, and the serving path that still runs on a Tuesday morning when nobody is watching it.
Where I did this
HKT, Group CTO Office
Delivered end to end as team lead: sourcing and cleansing, feature engineering, the model, and the serving path over Hadoop, Hive, Spark and MongoDB. Alongside credit analytics and spatial work with geohash.
End to end meant owning the pipeline, not just the model.
Not a smart assistant for the whole company, but one workflow end to end with nobody in the middle. A slice small enough to finish in weeks, deep enough that the business sees it. Whatever cannot show value in that window was not the core value.
Where I did this
Crewio, current
A scattered messaging estate consolidated behind one agent platform, with operational agents that act inside the systems of record rather than reporting on them: procurement chasing, usage analysis, and an approval step recorded before anything applies.
An agent may propose. A human applies.
Four areas, rather than everything I have touched.
Embedded with the customer rather than shipping product from a distance: finding the problem worth solving, proving it on their own data, and getting one workflow live before widening it. Product ownership and cross-functional leadership up to group level.
Orchestrating language models for performance and scale, and agents that act inside the systems of record rather than reporting on them, under a recorded approval step.
Ontologies, entity resolution, hybrid retrieval and provenance. Shared engines against per-customer deployments, and deciding what shares a database and what must not.
Recommendation engines, credit analytics, segmentation, OCR and NLP, over real operational data at group scale. Knowing what a messy production table actually looks like is why the real-data rule above is a habit rather than a slogan.
Jul 2026 — nowLeading engineering on an agent and knowledge platform for enterprise and SME customers: shared engines in a monorepo, per-customer deployments across three tenancy tiers, and the architecture underneath them. Where customisation is allowed to live, what a brain may assume, and what activation actually costs per customer.
Mar 2022 — Jul 2026Led development of AI-driven products from conception to deployment, with a focus on orchestrating large language models for performance and scale. Product owner of the document intelligence service and lead for AI development at group level; product manager across several data products. Rebuilt the business analytics pipeline with management in each business unit.
Sep 2022 — nowFounding committee member of Hong Kong's data practitioner community.
Aug 2020 — Mar 2022Group-level strategy and analysis across mall management, SmarTone, YATA, KMB and The Point. Five data sources fused into hundreds of customer segments, plus bus maintenance optimisation and OCR over vehicles and receipts.
Dec 2016 — Aug 2020Team lead. A recommendation engine delivered end to end, credit analytics, and machine learning pipelines over Hadoop, Hive, Spark and MongoDB.
Big data platform internship at SAP, and before that commercial and retail banking at Hang Seng Bank in Macau. The arc from banking operations to engineering took ten years.
Eric Tai. Forward-deployed engineer, Hong Kong.
Set in Inter. The garden is 1,433 points, and so is the code it becomes.