What we build
A personalisation layer that speaks for the brand — not a bolt-on that fragments it.
Each module is a production component — not a proof-of-concept — wired into your CDP, catalogue, and channel stack, with holdout-validated lift on every surface.
Recommendation engine across surfaces
Collaborative filtering + contextual bandits serve coherent recommendations on PDP, home rails, app, email, and in-store signage — same customer, same tone, every touch.
Personalised merchandising
PDP cross-sell, home-rail reorder, and category-page merchandising tuned per visitor — holdouts in place so lift is attributed cleanly to the personalisation layer.
Dynamic email + app content
Hero imagery, body copy, and product modules selected per segment at send time — winning variants promoted automatically through a managed bandit policy.
In-store signage feeds
Dwell-time signals and inventory-linked content feeds rotate signage per store, per hour — privacy-safe sensor counts only, no biometric capture anywhere on the floor.
RFV + wardrobe / fragrance modelling
Recency-frequency-value segmentation plus category-specific modelling — wardrobe continuity for fashion, fragrance-profile continuity for beauty, watch-collector cohorts for jewellery.
Ultra-luxury 1:1 curation
For the top-tier cohort, personalisation shifts to hand-curated lists reviewed by a stylist — the copilot drafts, the human edits, the client sees signature service.
Where personalisation earns its keep
Same layer. Different surfaces.
One feature store, one identity layer, one recommender — activated across every surface the brand owns. Category-specific modelling layered on top: wardrobe continuity for fashion, fragrance profile for beauty, occasion modifiers for jewellery.
Multi-channel retail brands
Brands running web, app, email, CRM, and physical stores — one unified personalisation layer across every surface, tuned per channel without fragmenting the profile.
Premium beauty · counters + D2C
Skin-routine continuity across visits, fragrance-profile matching, sampling-led nudges, and counter-linked signage — respecting the beauty consultation as a service moment.
Fashion · mass-affluent upwards
Wardrobe-based recommendations, capsule-launch orchestration, and look-led merchandising — with seasonal and occasion modifiers that align with the brand calendar.
Category pages + landing pages
Grid order and module copy tuned per visitor intent — search term, referral, and cohort signals blended with inventory for merchandise-right, conversion-right pages.
App home rails
Home-screen rails personalised per session — with bandit exploration so the system keeps learning without starving proven winners of distribution.
Lifecycle email + journeys
Behavioural and transactional streams personalised at send time — winback, post-purchase, replenishment, and VIP cadences running on the same model layer.
Model families we deploy
No single model covers every surface. So we ensemble.
Each model family covers a distinct personalisation problem — blending their outputs gives you coverage, resilience, and a consistent tone every surface can trust.
Matrix-factorisation, item-item similarity, and graph-embedding models blended into a single ranker — resilient to cold-start and tunable to each brand's catalogue shape.
Per-surface bandit layer explores new variants while protecting proven winners — exploration budget tunable per brand and per surface to balance learning and lift.
The classical retail segmentation, reframed for luxury and premium — value-weighted, cadence-aware, and refreshed daily so cohort routing stays current as clients move tiers.
Keeps recommendations coherent across web, app, email, and signage — the same customer sees a consistent message, never a contradiction between the email and the PDP.
Data sources wired into every decision
The signals that power coherent personalisation.
Pulled in parallel, normalised into a shared feature store, and respected by every consent and residency rule before any surface sees a recommendation.
Lift · attribution · provenance
A recommendation means nothing without the trail behind it.
Every decision is logged with its model, version, signals, and bandit policy. Every surface ships with holdout-validated lift. Every recommendation can be explained, audited, and — if the merchandiser says so — overridden.
- Holdout groups maintained per cohort
- Per-decision model + signal provenance
- Bandit policy version tracked
- Aligned to GDPR, PDPA, PIPL, CCPA, IAB TCF 2.2
Why Axccelerate for personalisation
Not a widget on the PDP.
A layer across the brand.
A widget gives you a carousel. A layer gives you coherent tone across web, app, email, and signage — with holdout-validated lift, consent propagation, and fairness reviews built in.
Pricing
Priced to the surface, not the impression count.
Personalisation deployments are custom — we scope against your surfaces, CDP, and channel mix before quoting.
Glossary
The vocabulary behind every personalised surface.
A quick reference for the terms that show up across merchandising, data, and compliance conversations.
- Collab filter
- Collaborative filtering
Recommendation approach that relies on behavioural similarity — 'customers like you also bought / viewed X' — without needing deep product-attribute knowledge. Strong for dense catalogues.
- Content filter
- Content-based filtering
Recommendation approach using product attributes — colour, category, price band, material. Reliable cold-start fallback when behavioural signal on an item is thin.
- Contextual bandit
- Online learning policy
An algorithm that picks between variants (creative, copy, product) while balancing exploration and exploitation — continuously learning which variant wins in which context.
- Cold-start
- Sparse-signal scenario
The case where a customer or a product has little history. Content filters and bandit exploration handle cold-start until enough behavioural signal accumulates.
- RFV
- Recency · Frequency · Value
A cohort segmentation method based on how recently, how often, and how much a customer has spent. Classical in retail, still the backbone of cadence and value-tier routing.
- Propensity score
- Action-likelihood estimate
A model's estimate of the probability a customer performs an action — buy, click, churn, upgrade. Propensities feed offer targeting, sequencing, and holdout construction.
- Cohort
- Behavioural / attribute segment
A group of customers sharing a defined behaviour or attribute — 'new SG app users, beauty first-purchase' — used as the unit of bandit arms, content variants, and analytics.
- Personalisation surface
- Where personalisation renders
A specific placement — PDP, home rail, email hero, signage panel — tuned independently. Surfaces share the same customer profile but carry different content contracts.
- Dynamic content
- Render-time customisation
Content (image, copy, product module) selected at render or send time based on the requester's profile and context — the mechanism that makes personalisation feel coherent.
- Lookalike
- Similarity-based audience build
An audience constructed by finding customers behaviourally similar to a seed group. Useful for acquisition, cold-start tuning, and product-discovery expansion.
- Segment of one
- Single-customer audience
The limit case of personalisation — a cohort size of one, with content tailored to the individual. Reserved for ultra-luxury 1:1 service and high-value clienteling motions.
- CDP
- Customer Data Platform
The identity, event, and consent layer that unifies customer data across sources. Every surface reads the CDP; every new decision writes back so other surfaces stay in sync.
- PII
- Personally Identifiable Information
Data that identifies a named individual — name, email, phone, address. Handled with residency, retention, and access controls aligned to GDPR, PDPA, PIPL, and CCPA.
- Consent
- Permission + purpose
The explicit grant to use customer data for a stated purpose — tracking, marketing, profiling. Recorded per purpose, per jurisdiction, and respected before any decision fires.
One customer. Every surface. Coherent tone.
30-minute scoping with a senior engineer and a retail-systems operator. You'll leave with a surface plan, CDP integration sketch, and realistic timeline — not a sales pitch.