I build AI-driven product and decisioning systems for regulated environments.

A UK retail bank with 500K+ customers and a £24B savings portfolio had 5 disconnected systems making identity decisions independently — document verification, fraud scoring, AML screening, consent capture, and authentication each running on different identifiers and thresholds. Customers got vague "we need more information" dead ends. Analysts spent their time stitching context across systems instead of catching fraud. Approved customers were later locked by systems using different rules.

I mapped the real decision path end-to-end (not the PowerPoint version), pulled p95 latency data by dependency, sat with analysts watching them work, and reframed the problem from "fix onboarding drop-offs" to "no one owns the customer's identity state across time." Designed an Identity Orchestration Service that normalized signals into a unified state model. Separated sync from async checks so only critical uncertainty blocked customers. Chose phased rollout over big-bang replacement — onboarding first, then monitoring, then cross-channel reuse.

No mobile banking product existed at this FCA-regulated UK bank. Legacy SOAP-based monolith on on-premises VMs. Batch-only wire transfers. Tightly coupled to an old core banking processor. Low daily active users, high customer churn, and a disconnected cross-channel experience. KYC and AML not integrated into any digital flow.

I led the cross-functional team (engineering, design, compliance, data science) to build the 0→1 mobile platform. Migrated to AWS ECS microservices with RESTful APIs and event-driven architecture. Integrated third-party KYC (Onfido), AML (ComplyAdvantage), and fraud detection (Sift). Designed progressive onboarding that cut steps from 10 to 4, with behavioral triggers for engagement. Used feature flags for phased rollout and A/B testing.

Rules-based fraud system protecting a £24B savings portfolio couldn't scale with 1M+ monthly transactions. Binary triggers flooded 100+ AML analysts with low-information cases. Reviewers acted as human middleware — stitching context across 5 disconnected systems to understand why each case was escalated. Manual review bottlenecks were growing every quarter.

I led the transition from rules to ML-driven risk models, working directly with the data science team to define model integration points, decision thresholds, and system workflows. Designed probability-score cutoffs replacing binary triggers. Built analyst-labeled feedback loops that fed back into continuous model retraining. A/B tested detection thresholds to optimize precision vs. customer friction.

100+ AML analysts struggling with binary fraud rules that triggered on simple thresholds. Cases arrived with no context — a senior analyst told me: "This one should never have reached me. The document passed, the sanction match is a false fuzzy match, and the device is new only because it's a new customer." Analysts were spending more time reconstructing narratives than catching fraud.

I analyzed p50/p95 latency data by stage, cross-referenced manual review outcomes with actual fraud capture rates, and sat with analysts for a week to see what they actually did. Redesigned the decisioning architecture from binary triggers to probability-score cutoffs with clear escalation tiers. Built structured case summaries so analysts got context upfront. Established feedback loops where analyst decisions fed model retraining.

SOAP-based monolith on on-prem VMs processing payments for a £24B portfolio. 2-3 patches per month consuming 25-30% of engineering bandwidth. Each patch broke other integrations — no backward compatibility, poor version control. Batch-only wire transfers while competitors offered real-time. Peak-hour incidents frequent and unpredictable.

I drove the architectural decision and phased execution. Rebuilt APIs on AWS ECS with containerized microservices and ISO 20022 messaging. Designed dual-path migration running legacy and modern APIs in parallel so 3 critical partners migrated at their own pace with zero downtime. Built sandbox environments for partner self-serve integration. Introduced async processing via Kinesis for reconciliation and confirmations.

Manual KYC and compliance workflows with a 15% error rate. Customer onboarding took 5 days. Audit failures due to inconsistent processes. Operations team spending 60%+ of their time on repetitive tasks with no clear audit trails for regulatory reviews.

I evaluated 3 RPA vendors head-to-head (Automation Anywhere, UiPath, Blue Prism), cost-benchmarked licensing, and ran pilot workflows before selecting Automation Anywhere for enterprise support and compliance controls. Deployed bots across KYC verification, account opening, document processing, compliance reporting, and fund transfer approvals. Built role-based access controls, structured retry logic, and audit logging for every bot action.

GDPR enforcement deadline May 2018. PSD2 deadline January 2018. 12 concurrent regulatory initiatives with hard deadlines and severe penalties. 5 competing stakeholder groups (product, engineering, compliance, legal, operations) each with different priorities. Policy language was unclear on technical requirements. Resource constraints across all engineering teams.

I risk-ranked all 12 initiatives by customer impact and regulatory penalty exposure. Scoped minimum viable compliance for each. Created shared specification documents translating policy language into functional engineering requirements. Ran bi-weekly risk-compliance syncs and trade-off workshops. Built steering dashboards for executive visibility and proactive blocker resolution. Tailored communication: technical deep-dives for engineering, business impact framing for executives, compliance assurance for legal.

250+ workflows across savings operations still using manual, paper-based processes. No disaster recovery plan for any workflow. Audit gaps in data handling. Poor process documentation. Staff unfamiliar with modern tooling. GDPR/PSD2 compliance gaps across all 250+ processes. Significant change resistance from teams comfortable with existing methods.

I led a cross-functional rollout team (Operations, IT, Compliance, Training) to upgrade the AWD platform with APIs connecting to CRM and core banking. Implemented automated validation rules and PSD2 strong customer authentication. Built data retention and deletion workflows for GDPR. Delivered phased upgrades by business unit to manage change resistance. Trained 100+ staff members and created SOPs and process documentation. Established the bank's first-ever disaster recovery capability for operational processes.

10+ business units (Finance, Operations, Compliance, HR, and others) each building custom automation independently. No shared process libraries or templates. Redundant vendor negotiations and tool selection across BUs. Inconsistent quality and compliance controls. Wasted resources on duplicate efforts that nobody was tracking at the enterprise level.

I defined the CoE charter and operating model, secured executive sponsorship and budget, and recruited automation champions from each business unit. Consolidated workflow patterns into reusable templates. Centralized vendor management into a single contract with better pricing. Built governance (steering committee + working groups) with quality standards. Mentored 3 junior PMs on automation scope and backlog grooming. Created a community of practice that sustained beyond the initial program.

~1,200 daily support inquiries overwhelming the team. CSAT baseline of 69 (bottom quartile). Limited personalization in customer interactions. Purely reactive service model with no proactive engagement. Fragmented journeys across mobile, web, phone, and branch. Long wait times and repeat contacts. Support team spending most of their time on basic inquiries that could have been self-served.

I built comprehensive customer journey maps in Salesforce, using behavioral data to identify high-friction drop-off points by segment. Correlated friction points with downstream churn patterns to prioritize interventions by impact. Deployed targeted help-layer content at friction points, contextual in-app guidance for complex flows, automated support nudges based on behavioral triggers, and proactive notifications for account status and pending actions. Measured impact by cohort and channel.

$50M portfolio spanning 8 concurrent platform, transformation, and infrastructure programs. High interdependencies between programs creating cascading delays. Execution delays and missed milestones before my intervention. Competing resource needs across programs. Hidden dependencies that nobody had mapped. Inconsistent reporting formats. Executive frustration with lack of transparency into what was actually at risk.

I built a RAID tracking framework with a portfolio-level dashboard showing health indicators for all 8 programs. Mapped cross-program dependencies to identify hidden cascading delay chains and enabled proactive mitigation before delays materialized. Ran capacity analysis, surfaced over-allocated teams, and executed targeted resource reallocation by strategic priority. Established weekly steering committee meetings, monthly business reviews with clear accountability, and defined escalation paths and decision rights.

45-day average loan processing time at an enterprise mortgage technology provider. Manual document collection and verification. Sequential approvals with no parallelization. Frequent rework due to incomplete applications and missing documents. Poor customer communication during the process. No visibility into application status for customers. CSAT declining at 68.

I conducted interviews with customers and loan officers, mapped current-state process flows, and identified unnecessary steps and handoffs. Removed redundant document requests and parallelized approval steps. Integrated the nCino lending platform with Experian automated credit scoring. Implemented document OCR and validation. Built a customer portal for real-time status tracking. Staged the redesign into risk-tiered releases with clean rollback at each milestone — zero service disruption across active loan applications during cutover. Target was 30 days.