NuStream formerly Libera · Retail Data Collection & Insight Platform
From photographs of store shelves to shelf-level insight for consumer-goods brands, across three countries.
01Overview
NuStream is a retail data-collection and insight platform operating across Indonesia, Pakistan and the Philippines. Field users submit photographs of store shelves and fridges; a computer-vision pipeline detects and identifies the products on display. A second pipeline ingests photographed receipts and extracts their transactions, keeping each store's data current.
The platform turns this raw field data into actionable insight for its customers, largely major consumer-goods suppliers and brands, who use it to understand product presence, distribution and sales at the shelf level.
02My contributions as Backend Engineer
Platform APIs
Built core service APIs for the platform: the admin-panel APIs used by store managers and regional administrators, and the Stores, Products and Submissions APIs that drive day-to-day operations.
- Services run on Google Cloud Run.
- Authored the GitHub Actions CI/CD pipelines for several of them (3–4 services), giving the team automated build-and-deploy.
LLM-based product auto-categorisation
Products entering a store had previously been categorised by hand, which produced inconsistent categories across the catalogue.
- Built a feature that assigns each product a consistent, three-level hierarchical category, modelled on the taxonomies used by established online product databases.
- From a product's name and title the system infers all three levels and writes them into the Neo4j product graph.
- The most granular level spans roughly 130 distinct categories, applied across tens of thousands of products. Accuracy was very high in practice, bounded mainly by how accurately product titles were entered.
03My contributions as Cloud Platform Engineer current
I own the platform's Google Cloud foundation, built entirely as infrastructure-as-code. Its defining choice is a cell-per-country architecture: each country is a self-contained "cell", its own GKE cluster plus a co-located data plane, linked to the others only through a shared management fleet. This keeps each country's data within its own region (data residency enforced per cell) and lets countries scale and fail independently.
Compute & network
- Private GKE Autopilot clusters on a Shared VPC, one cell per country (Indonesia is the primary cell).
- A consolidated management fleet via Connect Gateway.
Keyless, CI-only delivery
- All Terraform is applied exclusively from GitHub Actions through Workload Identity Federation.
- No service-account keys exist anywhere, and no changes are ever applied from a laptop.
Data & ML per cell
- Each cell carries its own Cloud SQL, Pub/Sub, object storage and a Qdrant vector store for product identification.
- Plus the ML foundation for the vision (GPU) and classification services.
Governance & observability
- Organisation-wide policies, IAM, centralised logging and audit/billing export.
- Structured in phases for auditability.
04Technology
05Impact
- Gave a multi-country product a secure, reproducible, residency-aware cloud foundation that scales country by country.
- Eliminated long-lived credentials from the whole organisation through keyless, CI-only delivery.
- Replaced inconsistent manual product categorisation with a uniform, automated three-level taxonomy across tens of thousands of products.
- Enabled the engineering team to ship services quickly on Cloud Run with automated CI/CD.
Architecture described at a conceptual level; internal identifiers and credentials are intentionally omitted.