We work under strict confidentiality. Client names, datasets and architectures are omitted; the work and the measured impact are not.
6 engagements
01 · Financial Services
Tier-1 Commercial Bank
-80%
Reconciliation cycle time
100%
On-premise execution
Autonomous AI Reconciliation Agent
Privacy-preserving autonomous agent that reconciles complex transaction ledgers entirely on-premises.
Context
A major bank needed to compress month-end reconciliation cycles across multiple ledgers without exposing sensitive transaction data to public cloud LLMs.
Approach
Designed and built an autonomous reconciliation agent on LangChain with privately hosted Llama 3 and Mistral models, grounded by a retrieval layer over the bank's internal rule corpus. The system ran fully behind the institution's security perimeter.
Outcome
Demonstrated potential to reduce reconciliation cycles by over 80% while improving reasoning accuracy through retrieval-grounded rule context — all within the organisation's security perimeter.
Default-prediction and risk-scoring system that materially reduced credit losses across a sizeable loan portfolio.
Context
The bank sought to reduce credit losses across a sizeable loan portfolio while maintaining responsible lending velocity.
Approach
Delivered a credit-risk decisioning solution combining gradient-boosted classification with risk scoring, deep feature engineering on borrower and behavioural data, and rigorous validation against historical outcomes.
Outcome
Loan-delinquency detection improved by 20%, with consistent model performance across validation windows and clear, auditable explainability outputs for the risk function.
Resilient ingestion and reconciliation platform that reshaped how the organisation manages transaction integrity.
Context
Reconciliation across multiple counterparties and rails was fragile, manual and lagging — eroding trust in daily transaction reporting.
Approach
Engineered a resilient ingestion and reconciliation platform with automated anomaly detection and analytics-engineering rigour to produce production-grade datasets.
Outcome
Total transactions reconciled consistently exceed 99.5%, materially improving accuracy, timeliness and transparency across the organisation.
Stack · Anomaly detection, analytics engineering, data quality
04 · Telecommunications
Regional Telco Operator
-50%
Subscriber churn
+30%
Retention uptake
Churn Prediction & Automated Retraining
Behavioural churn model with automated retraining that cut subscriber churn by half.
Context
Subscriber churn was eroding ARPU faster than acquisition could replace it, with limited visibility into early intent-to-leave signals.
Approach
Built a behavioural churn model on petabyte-scale event data with automated retraining and drift monitoring, plus an intervention recommender to route at-risk users to the right retention play.
Outcome
Subscriber churn reduced by 50% within twelve months, with the recommendation engine driving a +30% uptake in retention offers.
Single-source-of-truth warehouse and ML feature pipelines underpinning year-on-year revenue growth.
Context
Reporting was slow, error-prone and contradictory across geographies, blocking both operational decisions and downstream ML.
Approach
Engineered petabyte-scale ML feature pipelines and a single-source-of-truth warehouse, with rigorous lineage, quality controls and downstream BI integration.
Outcome
Reporting latency cut by 99%, data errors cut by 95%, NGN 50m+ in recovered revenue identified, and a +10% YoY revenue contribution from ML-driven decisions.
Stack · Spark, dbt, feature stores, MLOps
06 · Financial Services
Multinational Bank
1,000+
Staff trained
12mo
Programme duration
Enterprise Data Literacy Programme
Year-long, bank-wide data literacy initiative training 1,000+ staff in practical data usage.
Context
Front-line and middle-office staff lacked the shared data vocabulary needed to act on the bank's growing analytics estate.
Approach
Designed and led a year-long enterprise data literacy programme — curriculum, delivery, assessment and embedded coaching — across business and technology functions.
Outcome
Over 1,000 staff trained, measurable lift in self-service analytics adoption, and a durable internal community of data practitioners.