Background

At Standard Chartered, AI and machine learning models are used to process sensitive financial documents such as passports and cheques, helping reduce the manual verification effort involved in banking operations.

However, since the bank operates globally, these AI systems must comply with strict regulations defined by different countries and financial authorities. For example, Singapore follows the MAS FEAT guidelines for responsible AI, while European regulations such as GDPR require banks to maintain transparency, fairness, accountability, and auditability in their AI systems.

Although AI was introduced to reduce human effort, the bank was still spending a significant amount of manual effort on regulatory checks, audits, and compliance reviews. The existing validation process was highly manual, fragmented across multiple teams, and difficult to scale. Because of this, deploying a single AI/ML model could take anywhere between 6 to 9 months.

The biggest bottleneck existed in the final 20% of the model lifecycle, where compliance checks, audits, and approvals took place, but this stage alone contributed to nearly 80% of the overall deployment time.

To solve this, the bank developed FRIDAY (Framework Responsible for Intelligent Data and Algorithm Yield), a centralized platform designed to streamline the end-to-end AI/ML lifecycle.

FRIDAY automated critical processes such as model validation, bias checks, documentation, and monitoring, transforming black-box AI models into transparent and auditable systems. By reducing manual effort in the final compliance stage, the platform significantly accelerated model deployment and improved regulatory readiness across teams.

My Role


I was the Lead Product Designer for FRIDAY and was involved in the project from the BRD and SRS stages to ensure user needs were embedded into the requirements and workflows from the start. I also contributed to design strategy, stakeholder alignment, and overall project execution while collaborating with two other senior designers across different platform engines.

I defined the style guide and led the end-to-end design of the Data Suitability and Explainability engines, covering research, workflow mapping, service blueprinting, and UI design. This case study focuses on the Data Suitability Engine within FRIDAY.

Problem

Before a model could be trained, its data had to be validated for quality, fairness, and regulatory compliance across the countries where it would be deployed. This stage alone took 4 to 6 weeks per model, making it the most time-intensive part of the AI/ML lifecycle.

To better understand the causes behind these delays, I facilitated a collaborative workshop alongside the Product Owner with data scientists, project managers, and validation teams. Together, we mapped their workflows, identified the approximate time spent across each sub-step, and uncovered the key bottlenecks contributing to the overall delay.

Core Insight

The core bottleneck was the heavy dependence on fragmented manual workflows across validation, bias checks, PII masking, and compliance documentation, making the data suitability process slow, inconsistent, and difficult to scale.

Other Key Insights

• Fragmented tools and workflows across teams

• Manual compliance documentation and audit preparation

• Repetitive field-by-field PII verification

• No centralized audit trail

• Inconsistent quality validation standards

• Late-stage bias and fairness checks

• Lack of standardized documentation formats

• High dependency on manual cross-team coordination

Overall, the process was highly manual, inconsistent, and not scalable across key validation steps. This highlighted a clear opportunity to streamline and automate data suitability workflows, enabling the same level of rigor in significantly less time.

Solution


The solution was directly shaped by the research findings and operational bottlenecks uncovered during stakeholder workshops. The primary challenge was not model training itself, but the fragmented and highly manual compliance workflow surrounding data validation, fairness checks, PII verification, and audit preparation.

To address this, I collaborated closely with system architects, engineers, compliance stakeholders, and validation teams to identify:

• Which compliance activities could be automated reliably

• Which decisions still required human judgment

• How to standardize workflows without compromising regulatory rigor

Together, we established clear boundaries between system-driven automation and user-led validation tasks, ensuring the platform remained both efficient and trustworthy.

The result was the Data Suitability Engine, a centralized workflow within FRIDAY that transformed data validation from a disconnected, person-dependent process into a guided, transparent, and auditable experience.

The workflow was designed around five key stages:

1. Confirm data source

2. Perform automated quality checks

3. Conduct bias and representation audits

4. Validate PII masking

5. Generate a comprehensive suitability report

By structuring the process into a sequential workflow, teams gained better visibility, consistency, and traceability across the entire data suitability lifecycle.

Step 1 - Data source visibility and lineage

To address fragmented workflows and lack of audit visibility, I designed a centralized data source experience that gave validation teams a complete view of the training dataset in one place.

The interface surfaced key metadata such as source systems, regional distribution, document composition, ownership, extraction history, and lineage status. I also worked closely with engineers to ensure audit metadata and extraction logs were captured automatically instead of being manually maintained across teams.

This helped transform a fragmented and person-dependent process into a transparent and traceable workflow.

Solution impact: Reduced manual coordination and improved audit readiness.
Time saved: Eliminated manual data inventory tracking and reduced cross-team dependency.

Step 2 - Standardised automated quality checks

To reduce inconsistent validation practices, I designed a standardized quality check workflow that translated technical validation outputs into simple and actionable system feedback.

The interface consolidated checks such as missing values, duplicate detection, class imbalance, format consistency, image resolution, and outlier detection into a structured review experience. Status indicators, thresholds, warnings, and recommended actions helped users quickly identify issues requiring attention. This reduced the dependency on manual scripts and made validation results easier to review collaboratively.

Solution impact: Standardized validation workflows and reduced manual quality checks.
Time saved: Reduced days of manual validation work to automated checks completed within minutes.

Step 3 - Bias and representation audit

Fairness validation was previously handled late in the process, often causing delays during compliance reviews.

To address this, I designed an early-stage bias audit experience that evaluated dataset representation across dimensions such as geography, document type balance, customer segments, and time period coverage.

The interface surfaced FEAT compliance scores, warnings, and corrective recommendations directly within the workflow, helping teams identify representation gaps before model training began. My focus was on making fairness and representation risks easier to interpret for both technical and compliance stakeholders.

Solution impact: Enabled proactive fairness validation and reduced downstream compliance rework.
Time saved: Reduced late-stage rework by identifying representation gaps earlier in the lifecycle.

Step 4 - Guided PII masking with enforcement

To streamline repetitive privacy validation workflows, I designed a guided PII masking experience aligned with GDPR and FEAT requirements.

The interface displayed field-level masking decisions alongside masking techniques such as tokenization, redaction, generalization, and exclusion. Regulatory references and masking status were embedded directly into the workflow to improve transparency and auditability. This helped users clearly understand which fields required masking, why the action was required, and which fields could be retained safely.

Solution impact: Reduced repetitive field-level verification and improved traceability.
Time saved: Replaced repetitive masking and documentation workflows with a guided and auditable process.

Step 5 - Auto-generated suitability report

To reduce the heavy operational effort involved in audit preparation, I designed the final reporting experience to automatically compile validation outputs, bias audit results, compliance logs, and masking decisions into a centralized Suitability Report.

The report summarized overall suitability scores, flagged risks, compliance dimensions, and dataset metadata in a structured and reviewable format. Instead of manually consolidating information across systems, teams could now generate audit-ready documentation directly from the workflow. This transformed documentation from a separate manual process into an automated system output.

Solution impact: Reduced audit preparation effort and improved compliance readiness.
Time saved: Reduced compliance documentation from weeks of manual effort to minutes.

Impact

FRIDAY transformed the final compliance stage of the AI/ML lifecycle from a fragmented manual process into a scalable and audit-ready system.

Measurable Impact

• Model deployment timelines reduced from 6–9 months to under 30 days (projected)

• Data suitability validation reduced from 4–6 weeks to a few hours (projected)

• Addressed nearly 80% of deployment delays by optimizing the final compliance stage

• Reduced compliance documentation effort from weeks to minutes

• Improved standardization, traceability, and regulatory readiness across teams