Some of what we build starts as a question rather than a brief. This page is what we are currently working through, and where we think it is genuinely ready for a client's money rather than only interesting.
Language models, hosted and self-hosted
The useful question for a Malaysian business is rarely which model is cleverest. It is where the data is allowed to go, what it costs per month at your actual volume, and who is accountable when it is wrong.
On a provider's cloud. Fastest to start, strongest models, and no hardware to buy. The trade is that your text leaves your systems, pricing is per token and moves when the provider decides, and a model you built around can be retired on their schedule rather than yours. For most general work this is the right answer, and we say so.
Self-hosted, on your own server. An open-weight model such as Llama, Mistral or Qwen, running on hardware you control, is now genuinely practical for a narrow job. Nothing leaves the building, the cost is the machine rather than the meter, and the model does not change unless you change it.
We are testing where the line falls. Broadly, self-hosting starts to make sense when the task is repetitive and well defined, when volume is high enough that per-token pricing hurts, or when the data simply cannot leave. It rarely makes sense when the work is varied and open-ended, because a small local model will disappoint against a frontier one and no amount of tuning closes that gap.
Where it matters most. Legal, medical, financial and government-adjacent work all have confidentiality requirements that make a public API awkward. Manufacturing and engineering firms often have decades of specifications and drawings that they will not upload anywhere. Those are the cases we are building for.
Retrieval, which is the part that actually decides quality
A model on its own knows nothing about your business. The work that makes an assistant useful is retrieval: indexing your documents, finding the right passages for a question, and giving the model only those. Most disappointing AI projects fail here rather than at the model.
We are working on chunking that respects document structure instead of cutting every four hundred words, hybrid search that combines keyword and meaning, and answers that cite the document and page they came from so a person can check rather than trust.
Connecting to the major providers
We integrate with OpenAI, Anthropic, Google and the open-weight providers through their APIs, and increasingly through a single internal layer so that a system is not welded to one of them.
That layer handles the things that bite in production: switching provider without rewriting the application, falling back when one is rate limited or down, caching repeated questions so you are not billed twice for the same answer, capping spend per day, and logging every call with its cost so the monthly bill is explainable.
Model choice becomes a setting rather than a rebuild, which matters more than it sounds when prices and capabilities move every few months.
AI media for social content
Short-form video is where most small marketing budgets go, and the production cost is what limits it. We are testing generation and editing tools for the parts that are genuinely repetitive: cutting long footage into reels, generating captions and subtitles, producing variants of the same clip at different aspect ratios for each platform, voiceover in English, Malay and Chinese, and product images in consistent lighting without a studio booking.
Our position on this is unfashionable. Generated media is good enough for volume and variation, and not good enough to be the face of a brand. We use it to multiply work a person has already approved, not to replace the approval.
Anything generated should also be labelled as such. That is becoming a legal requirement in several markets and it is simply honest in the rest.
What this means for a client
We do not charge clients to be our test bed. When something on this page is ready, it appears in the services list with a price. Until then it is research, and we will say plainly which of the two you are being quoted for.
If your situation touches any of this, particularly if you have confidentiality requirements that rule out a public API, it is worth a conversation. Some of it we have already built.