Go back
Go back
CXP / Project Management
Restoring Control Under Delivery Pressure
Role
Product & Project Manager
Company
Feedloop AI - CXP
Type
Project Delivery
This project was part of a broader CRM transformation initiative aimed at strengthening Pegadaian’s data foundation, improving campaign effectiveness, and embedding machine learning capabilities into customer engagement workflows.
The scope included database architecture modernization, campaign performance and ROI enhancement, improved data integration, and the development of machine learning models to support customer profiling, segmentation, and Next Best Action use cases.
The objective was to evolve the CRM into a more analytics-driven and decision-oriented platform.
What i stepped into
I joined the program mid-execution.
At that point, the migration from ClickHouse to BigQuery had already been completed and was live in production. However, post-migration workload behavior did not align with earlier cost assumptions.
Within the observed period:
~661 TB of data was processed
Total cost exposure reached ~USD 4,100+
Daily BigQuery cost spiked to the IDR 55–60M rang
Rollback was not viable. The system was already in production. The priority shifted from questioning architecture decisions to stabilizing cost and restoring predictability.
Step 1: Establish Control Before Fixing
The first move wasn’t technical.
It was structural.
Before touching queries, I restructured how we approached the problem:
Separate cost stabilization from feature delivery
Protect critical ML and segmentation milestones
Avoid rollback unless absolutely necessary
Introduce a parallel workstream model
Instead of pausing everything to “fix BigQuery,” we split execution into:
Stabilization Workstream (cost & workload discipline)
Core Delivery Workstream (CRM & ML deliverables)
This prevented the cost issue from freezing the entire program.
That decision alone reduced delivery risk.
Step 2: Sequence the Recovery
Under pressure, sequencing matters more than speed.
We prioritized actions based on impact-to-effort ratio:
Identify highest TB-consuming queries
Reduce repeated full-table scans
Tighten partition filtering
Refactor heavy aggregation logic
Adjust evaluation frequency
This was surgical correction — not broad refactoring.
Each fix reduced cost exposure incrementally while keeping the system live.
Step 3: Manage Stakeholder Pressure
While engineering worked on stabilization, I focused on expectation management:
Transparent communication of cost drivers
Clear timeline recovery plan
Explicit trade-offs between speed and stability
Assurance that campaign operations would not degrade
Instead of promising “instant fixes,” we presented phased stabilization with measurable checkpoints.
Confidence returned gradually — not through slides, but through controlled execution.
The Result
Within weeks:
Daily cloud spend reduced from ~IDR 55–60M to below IDR 10M
Peak exposure reduced by 80%+
Monthly trajectory normalized (≈ IDR 1.3B equivalent reduction)
Campaign workflows remained stable
Delivery backlog brought back under control
The CRM transformation resumed with regained financial and operational discipline.
What This Project Actually Demonstrated
This wasn’t just query optimization.
It required:
Risk decomposition under live production pressure
Parallel workstream orchestration
Strategic sequencing of remediation actions
Balancing financial control with delivery velocity
Cross-functional alignment across Data, Engineering, CRM Ops, and Business
Most importantly:
The ability to restore control without escalating disruption.
