01 / FEATURE
Recommendations with cold-start
Hybrid recommenders that lean on content embeddings when there’s no behavioural signal — and switch to collaborative filtering as click data accumulates.
Personalisation at scale.
Relevance that survives catalogue sprawl, seasonal spikes, and the long tail. For retailers and marketplaces operating across ASEAN’s most fragmented consumer markets.
The pressure
Why retail AI underperforms.
Most retail AI we inherit was built on a tutorial and tested on the bestsellers. The long tail, the cold start, the seasonal cliff — that’s where the system breaks and the merchandiser stops trusting it.
Catalogue sprawl. Hundreds of thousands of SKUs across categories that don’t share attributes. Generic recommenders collapse to bestsellers and miss the long tail entirely.
Cold start, every season. New collections, new categories, new marketplaces — without months of click data, every recommender starts from zero. The seasonal cliff is the test.
Merchandising rules that matter. Margin guards, exclusivity windows, brand adjacency, inventory clearance — rules the recommender has to respect, not override in the name of CTR.
Multi-market, multi-language. Bahasa, Thai, Vietnamese, Tagalog — and the search queries customers actually type. English-tuned models miss intent at the door.
Where we ship
Use cases we’ve put into production.
Patterns we’ve shipped across ASEAN retailers and marketplaces — measured against revenue per session and margin per order, not vanity CTR.
02 / FEATURE
Search relevance
Tuned on your click and conversion logs, not a generic corpus. Multilingual query understanding for Bahasa, Thai, Vietnamese and the local dialects underneath.
03 / FEATURE
Merchandising co-pilots
Recommend collections, page layouts and promo placements that respect margin rules, exclusivity windows and inventory levels — not just predicted CTR.
04 / FEATURE
Vision-based shelf analytics
In-store cameras for footfall, dwell time, queue length and shelf availability — anonymised at the edge, PDPA-aware by design.
05 / FEATURE
Customer-service grounding
Order status, returns, sizing and product Q&A grounded in the actual catalogue and policy. Hands off to a human at the first whiff of complaint or refund.
06 / FEATURE
Demand and pricing intelligence
Forecasting that respects promotion calendars, weather, paydays and Lunar New Year. Pricing experiments with honest statistical power calculations.
Real-world example
A 22% lift on long-tail recovery.
A regional marketplace had a recommender that did fine on bestsellers and badly on the 80% of inventory that wasn’t. We rebuilt it as a hybrid with content embeddings for cold-start coverage and re-ranked against the merchandising rule set the category managers actually use.
Before
Status quo
- Recommender collapsed to top 5,000 SKUs
- Cold-start coverage under 20% on new collections
- Merchandiser overrode 30%+ of placements manually
- No multilingual query understanding
After
Post-rebuild
- +22% revenue per session on long-tail traffic
- Cold-start coverage at 75% on new collections within a week
- Merchandiser overrides under 8%; rule set encoded directly
- Search relevance lifted across Bahasa and Thai queries
ASEAN online marketplace · A/B test, 8 weeks
Solutions that fit
Where to start, by maturity.
AI Sprint — 4 weeks → Validate a recommender or pricing experiment end-to-end with a working prototype.
Accelerate — embedded → One senior engineer in your sprint cadence, 3–6 months, monthly cancel.
Generative AI → Customer-service grounding, product copy, merchandising co-pilots.
Vision AI — shelf → Edge-deployed shelf analytics with PDPA-aware anonymisation.
Compliance & assurance
Frameworks we build against.
Personalisation lives or dies by trust. We design with consumer protection and data privacy as defaults — not as a legal review at launch.
PDPA across ASEAN. Singapore PDPA, Malaysia PDPA, Indonesia PDP Law and Thailand PDPA — consent, retention and cross-border transfer rules implemented per market.
Consumer-protection rules. Honest discount displays, no dark patterns in the recommender, transparent ranking signals where the regulator requires them.
Vision in physical retail. Anonymisation at the edge, no biometric retention by default, signage and consent flows tested against local consumer rules.
Need a recommender that works on the long tail?
Book a 30-minute call with our retail lead. We’ll review the metrics that matter and recommend a starting tier.
Talk to retail lead
30 minutes · reply within 1 business day