01
E-commerce with 500+ SKUs where "people also bought" is still hardcoded.
Fits you
Rec audit · 45 min
+23%
Revenue lift
+18%
CTR increase
2.1×
Items per session
35%
Recommendation Engines
Amazon's recommendations drove 35% of their revenue. Yours could too.
Most "recommended for you" widgets are embarrassingly bad — showing the thing the user just bought. We build recommendation systems that actually move GMV.
E-commerceMediaEdtechFintechMarketplacesB2BE-commerceMediaEdtechFintechMarketplacesB2BE-commerceMediaEdtechFintechMarketplacesB2B
Who this is for
Who actually benefits.
If any of these sound like you, we should talk.
02
Media / OTT where watch-time is the primary business metric.
Fits you
03
Edtech where course sequencing and next-lesson suggestions drive completion.
Fits you
04
B2B marketplaces where supplier-buyer matching is currently manual.
Fits you
The proof
How a real recommendation happens.
User event
→Feature engineering
→Candidate generation
→Ranking model
→Business rules
→Rendered to user
What we build
Recommendation systems, by type.
01
Collaborative filtering
Users like you also liked…
02
Content-based
Similar to what you're viewing.
03
Hybrid + deep learning
Two-tower, transformer, graph.
04
Real-time personalisation
Session-based, in-flight adjustment.
ClientA D2C marketplaceE-commerce
<em>+23% revenue</em> from rebuilding "you might also like".
Replaced generic "similar items" with a proper two-tower recommender. Trained on 18 months of behavioural data. A/B-tested at 50% for 4 weeks.
01
23%
Revenue lift
02
18%
CTR increase
03
2.1x
Items per session
04
+₹0Cr/year
GMV from widget alone
What it's like working with us
We changed three widgets. Revenue went up 23%. Without spending another rupee on ads.
NA
Nandini A.
VP Growth · A D2C marketplace
E-commerce · Mumbai
Tech stack
The tools. Chosen for your reasons.
08technologies in rotation
01PyTorch
02TensorFlow Recommenders
03Vertex AI
04Redis
05Kafka
06ClickHouse
07Feast
08MLflow
Process
How we actually work.
06 stages
- 01
Goal definition
Week 1Optimise for clicks? GMV? Retention? Pick one.
- 02
Data pipeline
Weeks 2-5Events, features, training sets. The unsexy 60%.
- 03
Baseline model
Weeks 5-7Simple model, deployed, measured against current.
- 04
Iterate + A/B
Weeks 7-12Better models, A/B-tested, shipped if they win.
- 05
Productionise
Weeks 12-14Real-time serving, monitoring, retraining.
- 06
Ongoing
Week 14+Monthly retrains, quarterly architecture reviews.
Questions
Answers, without the fluff.
Still have questions? Talk to us — we answer within a business day.
07common questions
01How much data do we need?
3-6 months of event data is a reasonable starting point. Less than that, heuristics + content-based are better than ML.
02Real-time or batch?
Both. Batch for candidate set generation (nightly). Real-time for ranking (per request). Hybrid is the production norm.
03Cold start problem?
Handled via content-based fallback for new users / items, popularity for no-context cases, and explicit onboarding preferences where possible.
04How long does a typical engagement take?
Most projects run 10-18 weeks from kickoff to production launch. We share a milestone plan in week one and update weekly.
05Do you sign an NDA?
Yes. Standard mutual NDA on request, before the first technical conversation.
06Who owns the code and IP?
You do. Code is in your GitHub org from day one. All IP transfers unambiguously on delivery.
07What does your pricing model look like?
For v1 builds: fixed scope, fixed milestones. For ongoing work: monthly retainer with a defined team. We don't do time-and-material surprise billing.
Related
Explore more of what we do.
Recommendations that actually recommend.
A 45-minute audit of your current widget performance. We'll estimate the lift.
Free rec audit.