Services
// Data Platform & AI Engineering
Data your business can act on. AI that survives production.
This is the practice Digithun was built on — and the reason Google Cloud appointed us as 1 of 8 Agentic AI consulting and implementation partners in Thailand, alongside Accenture, Deloitte and NTT DATA. We build the full arc: data platforms that can be trusted, analytics people actually use, and AI that is still running when the demo season is over.
Talk about data & AI Request a current-state review
Google Cloud Partner & Authorized Reseller · Data Analytics Specialization · Healthcare & Life Sciences Recognition · 1 of 8 PanyaThAI Agentic AI Partners · ISO/IEC 29110 (SGS)
// Where the decision stalls
The dashboard exists. The decision still doesn't.
| Today | With a platform underneath |
|---|---|
| Three teams report three different revenue numbers, and the meeting becomes a debate about the data | One definition per metric, computed once, used everywhere |
| Customer data sits in the shop system, the CRM, the app and an agency's tool — and nowhere together | One customer profile, resolved across sources, that marketing can query without filing a ticket |
| Building an audience takes two weeks and an engineer | Marketing builds the segment and pushes it to the channel the same day |
| Every AI idea stalls at "where would the data even come from?" | A foundation that new use cases plug into instead of starting from zero |
// What we deliver
The full arc, in one team.
-
Data platform and warehouse engineering
Ingestion from your operational systems, a modelled warehouse on BigQuery, and transformation pipelines that run on a schedule instead of on someone's laptop.
-
Pipelines, governance and quality
Each metric defined once, lineage you can trace, and quality checks that fail loudly — so when finance and marketing disagree, the argument is about the business, not the query.
-
Business intelligence and self-service analytics
Looker models for the questions people ask every week, under our Data Analytics Specialization — and self-service for the new ones.
-
Customer Data Platform and single view
Identity resolution across channels, one profile per customer, and segments pushed into the systems that actually spend the budget.
-
GenAI applications
Semantic search, RAG and document intelligence, grounded in your own content rather than the model's imagination.
-
Agentic AI
Multi-agent systems and function-calling architectures on Vertex AI — designed, evaluated and handed over so your team can run them.
-
Recommendation and personalization ML
Trained on your behavioural data, deployed where the customer sees them — see Retail & E-Commerce for where this pays off.
-
Healthcare AI
Under our Healthcare & Life Sciences Recognition — governance first, then the use case, inside PDPA and patient-data constraints.
StackBigQueryVertex AIGeminiLookerVertex AI Agent Builderdbt
// How a project runs
Four phases. Each one leaves something you would keep.
-
Audit and definitions
We map every source, find where the numbers diverge, and agree the definitions that matter with the people who argue about them.
- A source inventory
- A data quality assessment
- An agreed metric dictionary
- A target architecture
-
Platform build
Ingestion, modelling and transformation on BigQuery, with access control and cost controls from the start — on a landing zone done properly.
- A working warehouse
- Documented models
- Pipeline monitoring
- A cost baseline
-
Analytics and CDP
Looker models and dashboards for the recurring questions; identity resolution and segments where a CDP is in scope.
- Production dashboards
- A self-service model
- Unified customer profiles
-
AI on top
The use case the platform was built for — GenAI, recommendations, or an agent with real tools and real guardrails.
- A production AI system with an evaluation set, monitoring and a runbook — and someone to run it, if you want that to be us
Where you start depends on what already exists. Most clients do not need all four.
// Where this comes from
Where this discipline comes from.
ShopSCAPE and Pantip MALL — the commerce platforms we build and operate ourselves — run on data platforms we designed: the same warehouse discipline, the same pipelines, the same recommendation systems we build for clients. When something breaks at scale, it breaks on us first.
Client names and figures appear here only after the client has approved them.
// Common questions
What teams ask before they start.
Do we need a customer data platform, or is a data warehouse enough?
If your questions are "what happened last month" and "which channel performed", a warehouse and good reporting will do. You need a CDP when you have to act on an individual customer — resolve them across channels, build a live segment, and push it into a tool that spends money. Many teams buy a CDP when what they actually lacked was agreed metric definitions. We will tell you which one you are.
Do we have to build the platform before we can do any AI?
Not always. A well-scoped agent or search use case can run against a narrow, well-understood set of documents without a full platform behind it. What you cannot skip is knowing where the source of truth lives and who is allowed to see it. If that is unclear, the platform work comes first — and we sequence it so the AI use case is what the platform is built towards, not an afterthought.
We already have Power BI / Tableau / an agency dashboard. Do we throw it away?
Usually not. The problem is rarely the reporting tool — it is that the numbers behind it are computed in four places. We fix the layer underneath and keep the tool your people already know, unless there is a real reason to move.
How long before we see something useful?
The first useful output is almost always a small set of trusted numbers, not a full platform — and that comes early. We sequence the work so each phase leaves something you would keep even if you stopped there.
How do you handle PDPA and customer consent?
Consent state is part of the customer profile, not a spreadsheet next to it. Retention, access rules and data lineage are designed in during the audit phase, so you can answer a data subject request without an archaeology project. For patient data, Healthcare AI carries stricter controls.
Can this run somewhere other than Google Cloud?
Yes. Most of our data platform work runs on Google Cloud, and we also deliver on BytePlus and Alibaba Cloud where the region, the workload or the commercial terms make more sense — with billing in Thai baht either way.
Start with the numbers you already argue about.
Bring us one metric your teams cannot agree on. It is the fastest way to see what is actually wrong underneath — and it is usually not the dashboard.
Innovate for the better tomorrow.
// Corporate update
Our
Move
-
Buying the platform and building on it used to be two conversations with two suppliers. It is one conversation now: we resell Google Cloud, Alibaba Cloud and BytePlus, and the same engineers who size the environment stay with it through production support.
For teams already running with us, nothing changes technically — the difference is commercial. Licensing, quota and billing sit with the people who know what the workload actually does.
-
A certified implementation partner works to the cloud provider's published reference architectures. In practice that means your landing zone, IAM model and network layout look like something any Google Cloud engineer can pick up — including the next team you hire.
It also means the review gates are not ours to waive. Where the reference architecture asks for separation of duties or a break-glass path, it gets built.
-
Hospital data arrives in fragments — HIS exports, lab feeds, scanned forms, free-text notes in Thai and English. Before a model sees any of it, someone has to answer where each field came from, who consented to what, and which records must never leave the country.
We build that layer first. It is slower to demo and it is the reason the pilots survive contact with a real ward.
-
Live commerce moves fast enough that the recommendation loop has to close in the same session. That puts the weight on the event pipeline, not the model: what counts as a view, when a cart event lands, how quickly the feature store sees it.
We treat the BytePlus components as a stack to be wired properly rather than a switch to be flipped. The lift comes from the wiring.
-
The site you are reading ships as static HTML, one stylesheet and one script, served by a Node process with a strict content security policy. There is no analytics tag, no font CDN and no tracker.
It is partly a statement of taste and partly a working sample: the same restraint we bring to a client's platform, applied to our own front door.
-
Warehouses fill up faster than they get governed. By the time a model needs a feature, nobody can say which of the four revenue columns is authoritative, and the project stalls in a meeting about definitions.
The fix is unglamorous: contracts on the ingest side, lineage through the transformations, and one owner per domain. Do that and the AI work stops being archaeology.