hải nam · Senior AI Platform Engineer

Projects

What I have built.

AI systems owned end to end.

Clients include a financial regulator, a medicines agency and a national postal operator. I take projects from scoping to deployment, lead small teams and mentor juniors. The common thread is the part after the demo: real users, other people's data and a system the client has to keep running. Most of it sits on Kubernetes with the cloud declared from inside the cluster, so an environment can be rebuilt from its repository.

Voice-agent platform

Callers ring a phone number and talk to a realtime speech model that books appointments, answers from the organisation's documents and hands over to staff. I started it and lead the small team behind it.

Cloud platform

I designed and built an Azure platform to carry every product that comes next: GitOps-managed Kubernetes with every cloud resource declared from inside the cluster. Passwordless and EU-resident by construction.

Meeting-minutes SaaS

Meetings, live or recorded, become formal minutes. I started it and led the backend and ML: streaming speech recognition, speaker diarization, a call-analysis API for a telephony partner and the Kubernetes operator that lets it run on a sovereign host.

On-premise assistants

Two assistants on a client's own GPU server, no external AI service: one critiques draft committee papers, one answers from the decision archive with citations. Open-weight models on vLLM; edits proposed as small, checked changes. I led the AI engineering.

Permission-aware assistant

Staff get answers from their organisation's own documents, with citations. I locked it down on Azure and guided the juniors who built its permission-aware retrieval, which only returns what the user may open and fails closed.

Email triage

Routes an organisation's incoming mail from the shared inbox to the right department; the uncertain rest goes to people. I owned the ML: a fine-tuned transformer on KServe, retrained and swapped in without a redeploy.

Delivery-date prediction

A survival model predicts a parcel's delivery time from its latest tracking update. I took it over, redesigned the model and built the monitoring that checks it against actual deliveries.

Research · POSYDON · 2018–2020

A framework for binary stars.

A core developer, from the project's first year, of a binary-star population-synthesis framework built by an international collaboration at the Université de Genève and Northwestern University.

Worked on the code that carries a binary between two detailed simulations: it matches the star onto a grid of precomputed stellar tracks by mass and core composition, then integrates the orbit until the star overflows its Roche lobe.
Helped build the binary and star classes, the hand-off from one phase to the next and the shared physics utilities that every simulated binary runs through, much of it still in the released code.
Master's thesis: a hybrid method that pairs a fast approximate code with detailed precomputed simulations; without tuning, its predicted X-ray luminosity functions matched observed star-forming galaxies within an order of magnitude.

posydon.org

Publications

Papers published with the collaboration, including the two that describe the framework and its first results. The software is the contribution; the papers are where it was used.

Fragos et al. 2023ApJS 264, 45The framework paper
Bavera et al. 2021A&A 647, A153Gravitational-wave populations

Full list on Google Scholar

Experience and education

Physics first, then production.

2023–present

Senior AI Platform EngineerDroids Agency, Copenhagen

Started the meeting-minutes product and led its backend; lead the voice-agent team and designed the shared Azure platform; designed a client's on-premise assistants and was the senior engineer on a permission-aware assistant.

2021–2023

Data ScientistDroids Agency, Copenhagen

Took over three inherited ML services and kept them running; redesigned the parcel delivery model as a survival model; built the email-triage classifier, a fine-tuned transformer, with the training, serving and retraining loop around it.

2018–2020

Software and numerical modellingPOSYDON, Université de Genève and Northwestern University

Core developer on a population-synthesis framework from its first year; simulation modules and interpolation along precomputed stellar grids.

2015–2020

M.Sc. PhysicsNiels Bohr Institute, University of Copenhagen

Computing, simulation and modelling. Thesis on the formation and evolution of X-ray binary populations, written inside the POSYDON collaboration and supervised by Tassos Fragos.

Skills

What I reach for.

Scientific computing

ML and AI

Systems

Identity and security

Azure

Programming

Languages

About

I build machine-learning systems that have to run: the model, the serving, the Kubernetes under it and the Azure around it. I came to it from physics, where I helped build a simulation framework that a whole collaboration still uses.

If you are building something that has to keep running, from the model to the cluster under it, I would like to hear about it.

Based in: Copenhagen, Denmark
Training: M.Sc. Physics, Niels Bohr Institute
Languages: Danish, English, Vietnamese

Let's talk.

The quickest way to reach me is the chat. The slowest is email, and it still works.