Client Accounts
What it is like to work with us, in their words
We asked clients to describe their experience with specificity, not just sentiment. These accounts reflect the actual engagements, including the parts that required adjustment.
← Back to HomeReviews
Client testimonials
Head of Data, Petaling Jaya
We had built a document classifier that performed well on our test set but kept failing in production in ways we could not predict. The production engineering engagement identified three distinct failure modes within the first week. The monitoring infrastructure they built has since caught two data pipeline issues before they reached the model. The documentation is thorough enough that our own engineers can now extend it without needing us to call them back.
April 2025 · Production Engineering
Engineering Manager, Kuala Lumpur
Our internal knowledge base search had deteriorated as the document volume grew. The retrieval engagement was more structured than I expected — they established exactly what "good search" meant for our users before changing anything, which meant we could see clearly what improved and what did not. The post-launch tuning round resolved the ranking issues that appeared with real user traffic patterns. One caveat: the engagement scope was tighter than I initially wanted, but in hindsight that was probably right.
March 2025 · Retrieval Re-engineering
Senior Data Scientist, Shah Alam
The weekly mentorship sessions are the most useful technical input our team gets. It is not a lecture — they review what we actually built that week and ask questions about our reasoning. That has been more effective for developing judgment than any training course. The quarterly written reviews are surprisingly detailed and have been useful for framing conversations with management about the team's progress.
Ongoing since October 2024 · Engineering Mentorship
CTO, Cyberjaya
We came in wanting a broader engagement than the production engineering scope allows. They were direct about that and helped us understand what we actually needed versus what we thought we needed. The eight-week engagement addressed the retraining pipeline, which was the actual bottleneck. It would have been easy for them to agree to more work; that they did not was reassuring.
February 2025 · Production Engineering
Product Lead, Bangsar
Our search corpus is mixed Malay, English, and some Mandarin content, which made previous retrieval vendors cautious about promises. Wira Intelligence had clearly worked with this kind of corpus before. The chunking approach they used was adapted specifically for our document structure, and the post-launch round addressed a query pattern we had not anticipated. The evaluation methodology they delivered is now part of our regular QA process.
April 2025 · Retrieval Re-engineering
Team Lead, Subang Jaya
Six months into the mentorship arrangement. The sessions are structured around what we are actually building, which is the right approach for a team that is learning on the job. The reviewer asks questions that expose gaps in reasoning I was not aware of. It has been useful even in weeks where nothing went wrong — the habit of articulating decisions clearly has improved how the team works together.
Ongoing since November 2024 · Engineering Mentorship
Case Studies
Three engagements in detail
Challenge
Fintech firm, KL: ML model degrading in production
A Kuala Lumpur financial technology company had built a credit risk scoring model that performed well during development but produced increasingly erratic outputs three months after deployment. There was no monitoring in place to detect when performance changed, and no clear path to understanding what had caused the degradation.
Engagement
The production engineering engagement identified that the model had been trained on pre-pandemic transaction data and was encountering a shifted distribution in post-pandemic spending patterns. Beyond the data issue, there was no evaluation pipeline that would have detected this. The engagement delivered: a monitoring system tracking feature distribution and output confidence, a retraining schedule with explicit evaluation gates, and documentation of the known data sensitivity areas.
Results
7 weeks
Engagement duration
2 incidents
Caught by new monitoring before reaching production
"The monitoring alone paid for the engagement within the first month." — Head of Technology
Challenge
Legal services firm, Selangor: internal knowledge search returning unhelpful results
A legal services organisation had accumulated several years of case documents, precedents, and internal guidance across three languages. The search system, originally built as a simple keyword search, was producing results that legal staff described as "technically present but practically useless." Staff had begun maintaining personal document shortcuts outside the system.
Engagement
The retrieval engagement began by establishing what "useful results" actually meant for three distinct query types common in the firm's practice. The corpus was re-chunked to respect document structure rather than splitting by character count, which was the prior approach. Embedding model selection was adjusted for the multilingual content. Post-retrieval ranking was introduced to surface document recency and document type relevance. A post-launch tuning round addressed query patterns that appeared only after full staff rollout.
Results
68%
Improvement in retrieval accuracy on evaluated query set
Eliminated
Personal document shortcut workarounds within 6 weeks of launch
"The measurement methodology was as valuable as the improvement itself." — Operations Director
Challenge
E-commerce company, Johor: data team taking on AI responsibilities without specialist support
A Johor-based e-commerce firm had hired a data team with strong SQL and analytics skills to build AI features for their platform. The team had begun using ML frameworks but had no senior engineer with production AI experience to check their work. Mistakes were being discovered in production rather than during development.
Engagement
The mentorship arrangement began with a baseline assessment of the team's current work. The first four sessions focused on evaluation methodology — the team was making model decisions without clear success criteria. Subsequent sessions reviewed recommendation and ranking systems the team was building, with emphasis on identifying failure modes before deployment. By the third month, the team was raising failure mode questions themselves before implementation rather than after.
Results
0
Production incidents from AI features after month 3
Renewed
Arrangement renewed for a second six-month term
"The team is now asking the right questions before we build, not after." — Engineering Lead
Contact
Reach us directly
Credentials
Registered AI services provider, Malaysia
Technology company certification
Personal Data Protection Act adherence