Independent AI practice · one-north, Singapore

Why this is a practice and not an agency

I started Singaporebuildsai in 2026 because I wanted the person writing the evaluation set to be the same person on the call. The room is small on purpose.

How the practice began

In 2026 I opened this practice after watching teams in Singapore buy an “AI platform” and then spend a quarter trying to feed it a workflow the platform did not understand. The purchase came with slides, a named success manager, and a sandbox that never quite saw production data.

What those teams needed was narrower. A retrieval path over the documents the desk already trusts. A parser with an exception queue. A forecast whose error is reported in units operations already counts. Someone who would say when a two-page rules change would do the job.

A practice with one engineer can keep that promise. I write the code, I sit with the examples, and I stop when the harness says the model is not helping. There is no utilisation target that needs filling with a second workstream.

The office is in one-north, a short walk from the MRT. Most weeks I am at the desk; some weeks I am on a call walking through a failing test with the people who own the workflow. The geography is part of the offer: when data must stay in Singapore, so do I.

Principles

Method

Evaluation first

We agree on examples of a good output before we pick a model. The test set is versioned with the code. If we cannot write twenty honest cases, we are still in scoping, and that is a cheaper place to pause.

Models

Smallest model that passes

I start with the smallest open-weight or hosted model that clears the harness. Larger models are a last resort, because they cost more per call and are harder to keep inside a residency boundary. Passing the set is the only promotion criterion.

Data

Data stays where it must

If records cannot leave Singapore, the architecture follows that rule. Hosted APIs are used only when you permit them. I would rather have a slower local path than a clever path that ships personal data to a region you have not approved.

Delivery

Handover is the product

A working demo in my environment is unfinished work. You own the repository, the runbook, the test set, and the two people who can operate it. Until those four things exist, I have not delivered.

The stack I reach for

The list below is a toolkit. Your constraints pick the tools: where the data may live, who will operate the system after I leave, and how often the index or the model may change. I would rather inherit a Postgres instance your team already runs than introduce a fashionable store nobody can restore from backup.

Python and FastAPI cover most services I ship. Postgres with pgvector is enough for many retrieval stores; OpenSearch is there when the corpus and the query mix need it. Small open-weight models run when residency or cost requires it. Hosted model APIs are used when you have already accepted that path. Scheduling is Airflow if you have it, cron if you do not. Evaluation runners are pytest, because your engineers can read a failing test without learning a new product.

I document the why for each choice in the repository README, including the options I rejected. That note is for the person who will change this in a year, which may be someone I have never met.

Python FastAPI Postgres + pgvector OpenSearch Small open-weight models Hosted model APIs when permitted Airflow or cron pytest-based eval runners

What I do not do

I do not sell licences, platforms, or seats. You keep the code. If a hosted model is in the path, you hold that account, and I work inside it.

I do not take programmes whose main deliverable is a hundred slides on AI strategy. If you need a written view of one workflow, that can sit inside a discovery sprint. Strategy without a test set is a conversation I will not bill as a build.

I do not run several builds in parallel. When I am in a harness week, that is the week. If I am already engaged, I will say so and give you a realistic start date rather than a courtesy yes.

I do not quote accuracy without a test set drawn from your examples. A number from a vendor card, or from a public leaderboard, is not a promise about your documents. We measure on your cases, or we do not measure.

The room I work from

A quiet desk in one-north, a small table for scoping conversations, and a corridor I walk when a failing test needs a second look.

If this shape of practice fits

Write a short brief about the workflow. I will tell you whether I can take it, and in which format, within one working day.

Start a brief