Detailed Case Exposition
Business Outcome & Strategic Leverage
The relevance uplift translated directly into longer sessions and higher click-through revenues, validating Gibiru’s premium, privacy-centric positioning. The decoupled architecture now serves as a platform for future vertical search and advertising modules without vendor lock-in.
1 · Strategic Context & Market Friction
Privacy-oriented search engines often lag in result quality because they avoid behavioural data. Gibiru’s monolith compounded the gap—static term-frequency logic, no semantic understanding, and zero experimentation capacity. Competitive pressure from mainstream engines demanded a relevance leap without sacrificing user privacy.
2 · Objectives & Delivery Constraints
- Mandate: Ship an MVP in ten weeks that lifts precision by at least 25 %.
- Constraints: Self-hosted on GCP; no managed SaaS search; six-person engineering pod; strict privacy guardrails.
- Trade-offs: Limit BERT inference cost by re-ranking only top-10 BM25 results.
3 · Technical Architecture & Infrastructure Decisions
| Layer | Decision | Rationale |
| Query Processing | Elasticsearch + BM25 | Immediate lexical precision and flexible query DSL |
| Semantic Re-ranking | bert-base-uncased, cosine similarity on CLS token | Lightweight, domain-agnostic embeddings |
| Weighting | 0.6 × BERT + 0.4 × BM25 | Empirically maximised P@10 without latency spikes |
| Orchestration | Docker micro-services on GKE | Decouple modules; enable blue-green experiments |
| Feedback Loop | Clickstream → Pub/Sub → weekly retrain | Continuous learning without personal user IDs |
| Storage & Logs | Cloud Storage for artefacts; Stackdriver for metrics | Cost-effective, privacy-compliant observability |
Latency, scalability, and security were evaluated against a 95th-percentile query budget of 250 ms.
4 · Implementation & System Workflows
- Query → Elasticsearch returns top-10 BM25 hits.
- Re-ranker Service fetches cached document embeddings, computes BERT similarity, and fuses scores.
- Result API delivers ordered list to the Search UI.
- Clickstream Collector captures clicks, skips, and dwell time into Pub/Sub.
- Retraining Job aggregates pseudo-labels weekly, re-weights BERT, and updates Docker image via CI/CD.
Version control, container tags, and Canary deployments ensured zero downtime.
5 · User Experience & Product Storytelling
End-users experienced markedly sharper first-page results without perceiving architectural change—respecting the brand’s privacy ethos. Internal dashboards visualised P@10, NDCG, and dwell distributions, empowering product managers to communicate impact credibly.
6 · Performance Outcomes & Measurable Impact
| Metric | Pre-Overhaul | Post-Overhaul |
| Top-N Precision | Low-40 % | 65–70 % |
| CTR + Dwell | Baseline | +40 % |
| Experiment Cycle | ~30 days | < 7 days |
| Queries / Month | 10 M | 10 M + (stable) |
7 · Adoption & Market Strategy
The MVP shipped in ten weeks, followed by controlled A/B exposure to 10 % of traffic. Positive metrics triggered a full rollout. Isolation of ranking modules created a low-risk path for future semantic, multimedia, or RAG-based extensions.
8 · Feedback-Driven Evolution
Dashboard insights revealed query classes (e.g., medical topics) with residual noise. Targeted fine-tuning cycles and BM25 parameter sweeps closed precision gaps. The weekly retrain cadence established a culture of continuous relevance optimisation.
