AI, proven on our own work first. 

Our AI Innovation Lab is where we test AI on our own systems, measure it honestly and keep only what holds up. What passes becomes a solution you can use. What fails, we write down too.

Home › AI Innovation Lab

AI Innovation Lab · part of Luvina AX, our AI transformation programme

Bring us a problem

What you can use today

Most AI pilots never make it to real work.

Why a lab

In 2025, MIT found that 95% of enterprise generative AI pilots had no measurable impact on profit and loss.1 Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027.2 The gap is rarely the model. It is testing, integration, governance and running it after launch.

So we start from the work, not the tool: which process, which decision, which hours. We change how that work is done, measure it against how it was done before, and put the governance around it to run it every day. That is what we mean by AX, AI transformation.

We did it to ourselves first. Our own delivery runs on it: about 39% less effort across design, coding and testing. We call it LaunchEdge.

Sources: 1 MIT NANDA, The GenAI Divide: State of AI in Business 2025(opens in a new tab). 2 Gartner, June 2025(opens in a new tab).

AI pilot

AI transformation

A process, decision or cost you want to change

Starts with

A tool or a model

Success means

Owned by

After launch

A working demo

The IT team

The pilot ends

Hours, cost or errors down, measured against a baseline

The business, with IT

It runs every day, with governance and support

What you can use

What you can use today, and what is coming.

Each solution started as an experiment on our own work in the lab. The status shows how far it has come. We only quote numbers we have measured.

Live
Product companies · Corporates
LaunchEdge
Our AI-augmented way of building software
AI across design, coding and testing, with a senior engineer reviewing every AI-generated change and effort measured every sprint.
~39% less effort on average across our delivery. 51% on our own ERP.
Start:  on any project or dedicated team.
Pilot
Corporates
Legacy X-ray
Your old system, read and explained
AI reads your legacy code (COBOL, VB6, older Java and .NET) and our engineers turn it into a living specification, a dependency map and a migration risk report. Then you decide what to keep, rebuild or retire.

Legacy modernisation is one of our core strengths. AI reading of legacy code is on Trial in our verdicts.
Start:  fixed scope, [2-3 weeks, price]
In the lab
SMEs · Corporates
Agent Bento
Ready-made AI agents for everyday work
Invoices and purchase orders read straight into Dynamics 365 Business Central. Inbox triage with drafted replies. Document checks. A person approves anything that matters.
Start:  a 4-week pilot at a fixed price
Pilot
Product companies · Corporates
FDE
Forward Deployed Engineer: one problem, one engineer, four weeks
One of our engineers joins your team for 4 weeks to turn one business problem into a working AI prototype, then gives you a written recommendation: build, change course, or stop.
The same model took an EduTech client from a stuck prototype to a production MVP. Read the case →
Start:  4 weeks [price]
On the bench
Testing next
Reading engineering drawings into parts and BOMs for Aras PLM. Checking Dynamics 365 customisations before each Microsoft update. Agents that keep test suites current.
See what is on the bench →
Draft note: AX Center to confirm each status (Live / Pilot / In the lab), durations and prices. Check every name for trademark conflicts before launch. If a name changes, also update the links on the SME, Corporate and Product pages and the Solutions menu.

Adopt

Proven in our delivery. We use it by default.
AI support for test case design.
AI-assisted coding with mandatory human review
Proven in our delivery. We use it by default.

Trial

Working in selected projects. Measured closely.
Forward-deployed engineers for AI products
AI reading and documenting legacy code before migration
AI drafting of basic design documents

Assess

Promising. Still in the lab.
AI support for test case design.
Document intelligence for drawings and BOMs into ERP and PLM
Agents that run multi-step development tasks end to end

Hold

We advise against it today.
Measuring AI success by lines of code produced
Client data in public AI tools without data classification
AI-written code shipped without human review

What we would use, try, watch, and avoid.

Lab verdicts 

[Draft verdicts based on the experiment log and governance rules. AX Center to confirm each item before publishing. Offer "Lab notes" by email twice a year when the edition is released.]


Our current position on AI in software delivery, based on what we have tested ourselves. Updated twice a year.

What we have measured

Where AI actually saves effort - stage by stage

Measured across our own delivery. Productivity only; AI tool licensing and usage costs excluded.

Stage

Basic design

Detail design

Coding

Unit testing

Test cases

System and acceptance testing

Project management

Overall

Design support

AI-generated code, senior-reviewed

AI unit tests

Manual test support

Japan office established

Partially automated test support

Reporting and coordination

Across all stages

How AI is used

~29%

~15%

54-98%

50-98%

27-80%

30-50%

20-57%

~39% (16-51%)

Effort reduction

Experiment log

Numbered, measured, and reported - including what went wrong.

No

N.4

N.8

N.11

Status

What we learned

Our internal ERP, built from scratch with AI across design, coding and testing

What

EduTech product: a vibe-coded prototype rebuilt for production

Warehouse management system for a firm in Malaysia

Experiment

Production

Production

51% less effort; 187 to 90 man-months; ~6 months faster

Production MVP in 6 months after 12 months stuck

6 weeks vs an estimated 6.5 months; 72.5% lower cost

Result

Biggest gains in coding and unit testing (66%); requirements work saw none

A forward-deployed engineer closes the gap between business problem and AI architecture

Unplanned change requests added regression defects. Speed needs standard process and transparency

[N.x]

[Add other numbered experiments from the AX Center, N.1 to N.12]

Testing next

What is on the bench now.

Not proven yet. We will report results in the experiment log, good or bad.

Experiment

Keeping test suites current with agents

Agents that update tests when code changes, so regression coverage does not fall behind.

Experiment

AI review of ERP customisations

Checking Dynamics 365 extensions for upgrade risk before each Microsoft release.

Experiment

Drawings to product data

Reading engineering drawings into structured parts and BOMs for Aras PLM.

[Confirm the 3 current experiments with the AX Center.]

IP clause covering AI tool use in your contract
Mandatory senior-engineer review of AI-generated code
Safe by design
Data classification: what may enter AI tools, and what may not
Monthly KPIs: velocity, defect rate, AI code ratio
Audit trail of AI-assisted changes
Everything that leaves the lab carries the same rules.
AI governance in delivery →
Luvina Academy, developing engineering talent since 2008
Growing AI engineers
Delivery teams trained on AI methods
AI process standardised and enforced at team level
A growing pool of AI-ready engineers
Capability that scales without diluting quality.
AI platforms
The right model for each job, and partnerships on the way.
We build on leading AI models, chosen per project for cost, data location and fit, including Claude (Anthropic) and GPT (OpenAI, directly or through Microsoft Azure). We are now working towards formal partnerships with leading AI model providers, starting with certifying our engineers. [Confirm that certification has started]
Draft note - switching states: change data-state="B" on this block to A, B or C. A = nothing started. B = working towards partnerships (current; wording approved by Ethan). C = partnership signed. Anthropic calls a firm a partner only from its Select tier: 10 certified engineers, 2 joint customer deployments in production in the past 12 months, and 1 public customer story. OpenAI does not publish its tier thresholds. Delete the unused states at launch. Update the "AI platforms" row on the procurement page at the same time.

Lab with us

Bring us a problem worth experimenting on

We scope a short experiment on your problem, run it in the lab, and show you what works before you commit to building anything.

Path A
You have a pilot that is stuck
A prototype or AI demo that works on stage but not at scale. We assess it, then tell you what to keep, rebuild or drop.


Path B
You have a process that eats hours
Re-typing documents, forecasts in spreadsheets, manual checks. We qualify it and test whether AI actually helps.
01
Qualify
One call and a short review of your data and goal.
01
Qualify
One call and a short review of your data and goal.
01
Qualify
One call and a short review of your data and goal.

[Define the offer: price, data handling and IP of results.]

Lab sessions

Small roundtables for leaders figuring out AI.

One accountable team

A few times a year we host closed-door sessions in Singapore for business and technology leaders: what we tested, what failed, and what peers are seeing. No pitches.

At a session

[Confirm format, frequency and first date]

8 to 12 leaders, around one table
Results from our latest experiments
Open discussion under the Chatham
House Rule
A short write-up for attendees afterwards

Luvina FlexPool

A whole technology team, for less than one hire.

No. It is Luvina's own AI lab, part of Luvina AX, our group-wide AI transformation programme. The same teams that experiment there deliver client work, so what we learn goes straight into projects.
Is the lab a separate company?
LaunchEdge is live on client work now. Solutions marked Pilot are running with selected clients, and those marked In the lab are still being proven. Each card shows where it stands.
Can I buy the solutions today?
No. Client data stays in your environment and is classified before any AI tool is used. We follow your rules on what may leave it. [Confirm wording with ISO 27001 policy]
Do you use client data to train AI models?
[Confirm: client owns findings and code built on their data]
Who owns the results of an experiment?
[Price or price range for a 2 to 4 week experiment]


What does a lab experiment cost?

No. Many good experiments start with a manual process and a question: could this be faster or more accurate?
Do I need to be using AI already?

Curious what AI could do for you?

Tell us the problem. If AI can change how that work gets done, we will show you how. If it cannot, we will say so.

Talk to us

See the solutions

Copyright © Luvina Software JSC 

Terms of Services | Privacy Policy

Driven by passion and innovation, we channel Vietnamese intellect into impactful global IT solutions.

About us

Solutions

Technologies

Case Studies

Resources

Blog

Our Offices

Vietnam

Japan

United States

Singapore

Contact us

info@luvina.net

sales.en@luvina.net

Contact for Sale: (+84) 888 739 363

Contact for Career: (+84) 865 032 899