↩ Work / EXPERIENCE

Data Director.Employment

Period: Apr 2025 - Jul 2026
Company: GOZEM ↗
Link: Gozem Data Hub ↗

Owned data platform engineering and internal data products (the Data Hub and 30+ automation tools) and co-led AI strategy across Gozem's footprint: five African operating countries with Singapore as headquarters.

Architected and deployed Gozem Data Hub, a centralized platform serving over 650 people (internal employees, field agents, and external partners) with self-service analytics, AI and automation products, and data-driven applications for operations and business decisions.
Built line-of-business and operations apps (driver inspection, invoice extraction, external-acquisition agent performance monitoring, custom dashboard publisher, and others) used daily across Gozem’s operations.
Built G-Force, a mobile-first web application giving external acquisition agents visibility into submissions, validation status, rejection reasons, and payments, with approximately twice-daily synchronization.
Co-shaped Gozem’s AI strategy and shipped 30+ production AI and automation tools spanning fraud alerts, data extraction, and market monitoring; introduced AI-assisted engineering workflows to leadership and product engineering.
Built observability across 250+ pipelines (Dataform, Airflow, cron jobs) with automated data-quality checks, discrepancy alerts catching incomplete data and duplicates, performance tracking on exceptional events, and a data-exploration MCP.

[•] PROJECTS & USE CASES

8 ENTRIES

Gozem AI Strategy knowledge base

Result 49 validated use cases catalogued; 6 in implementation, with near-full top-management adoption of daily AI assistance.

Problem / Need

Gozem needed a coherent AI strategy spanning operations, logistics, driver management, ecommerce, vehicle financing, customer experience, and fintech, backed by a structured view of the AI vendor and tooling landscape.

Solution

Documented internal AI needs across 9 business units (Transport, Ecommerce, Financing, Finance, Data, Product, Marketing, Operations, Sales) with 50+ stakeholders contributing. Conducted a structured market scan covering 133 tools and vendors across 38 categories: AI vendors, foundation model providers, and automation platforms relevant to Gozem’s stack. Catalogued 49 validated use cases. Built an adoption framework around the 6 moved into implementation: content creation and daily assistance (near-full adoption by top management), coding and code review (rolling out with the product team), classifier and others (in use by the data team). Each use case ships with its agentic workflow architecture.

Tools

Data collection and analysis·Internal stakeholders interviews·Market data scraping·LLM-powered tools profiling·Strategic framework documentation

Gozem Data Hub: Central data & AI workspace

Result Architected and built the platform, personally coding 17 of its initial 19 microservices; the governed workspace served 650+ employees, field agents, and partners across 5 countries.

Problem / Need

Business and operational teams alongside some external partners across Gozem’s countries lacked a unified workspace for data and AI applications with fast iteration.

Solution

Architected and built the Data Hub, personally coding 17 of its initial 19 microservices. Two engineers each coded one service; I deployed one and reviewed and fixed the other before deployment. Designed database schemas, persistence models, centralized authentication, routing, and scoped access controls. Beyond the initial microservices, documented the shared architecture and contribution process, then used live debugging to onboard engineers to service/frontend integration, CI/CD, and code review. The platform supported data-driven applications, operational workflows, automation, data exploration, and self-service analytics for people collaborating across Benin, Togo, Gabon, Cameroon, and Congo.

Tools

React·FastAPI·Nginx·MongoDB·Redis·BigQuery·Google Cloud Storage·NATS·Docker Compose·Artifact Registry·Shared Python utilities

Application-serving data pipelines

Result Delivered application-ready datasets for internal operational applications, services, and microservices, extending the platform beyond BI outputs.

Problem / Need

Operational tools required governed, application-ready data from systems with different storage patterns, not dashboard-only extracts.

Solution

Designed Airflow-orchestrated pipelines sourcing MongoDB, BigQuery, Google Cloud Storage, and Google Sheets. Modeled the database schemas and persistence layers for data-intensive services, including data generated by application use and operational workflows.

Tools

Apache Airflow·MongoDB·BigQuery·Google Cloud Storage·Google Sheets·Data modeling

Platform delivery and reliability controls

Result Established repeatable service delivery with health monitoring, recoverable job execution, security gates, staged images, and rollback to prior images.

Problem / Need

A multi-service operational platform needed consistent release and runtime controls without overstating test coverage or disaster-recovery maturity.

Solution

Implemented service health checks, dependency and restart controls, execution logs, configurable retries, and stale-execution recovery. Established GitLab CI/CD with linting, targeted pytest suites, security and container scans, multi-architecture image builds, Artifact Registry publication, staged promotion, and rollback to previously published images.

Tools

GitLab CI/CD·Docker·Artifact Registry·Pytest·Trivy·Prometheus

Governed KPI reporting & data-warehouse MCP

Result Business users gained a shared path to KPI reports built from documented metric definitions and canonical warehouse queries.

Problem / Need

Business users and AI-assisted workflows needed consistent KPI definitions, source tables, transformation logic, and query paths across the warehouse.

Solution

Documented validated metrics with their source tables, transformation logic, and canonical queries; exposed the data warehouse through an MCP server and reusable BigQuery/Dataform views consumed by Looker Studio. This gave business users a shared reporting path grounded in documented sources.

Tools

BigQuery·MCP·Python·Dataform·Looker Studio

Automated invoice data extraction

Result ~40 hours saved per financing manager per month across thousands of invoices.

Problem / Need

Financing Operations team manually entered data from supplier and operational invoices in multiple formats, slowing reconciliation across Gozem’s multi-country operations.

Solution

Prototyped the invoice-extraction use case in Vertex AI Studio, testing prompts and context before production development. Architected and built an automated pipeline ingesting thousands of invoices per month (PDF and image formats) using LLM-powered OCR and document intelligence. Designed an agentic workflow supporting custom schema extraction, with validation to flag anomalies, retry handling, and human review, integrated into Google Sheets for reconciliation. Saved an average of 40 hours per financing manager per month.

Tools

Vertex AI Studio (prototyping)·LLM OCR·Document Intelligence·Python·Cloud Storage·Google Sheets

RAG-powered customer support chatbot

Result Resolvable support requests handled autonomously; complex cases escalated to Zendesk with conversation summaries preserving context.

Problem / Need

Customer support faced rising request volumes, with human agents spending too much time on resolvable cases and losing context when complex cases needed escalation.

Solution

Oversaw development of a multi-agent Retrieval-Augmented Generation chatbot handling Gozem support requests autonomously, routing resolvable cases to the chatbot and escalating complex cases to Zendesk. Added automatic conversation summarization to preserve context for human handoff.

Tools

RAG·Vector DB·Python·Zendesk

Real-time market monitoring system

Result +10.2% supply rate, +7.4% request-to-completion, +5.4% coverage at the Vodun Days festival peak (~2,000 users).

Problem / Need

Transportation operations had no live visibility on supply-demand imbalances during high-demand events (public holidays, major urban events), leading to slow operational response.

Solution

Architected a minute-granularity pipeline aggregating GPS, booking, and driver-availability signals. Built supply-demand imbalance detection with threshold-based alerts, plus an operational dashboard surfacing demand-density maps, supply-gap indicators, driver engagement scores, and event KPIs. Implemented automated Google Chat alerts and validated during the Vodun Days festival peak: 10.2% improvement in supply rate, 7.4% increase in request-to-completion rate, and 5.4% coverage gain across ~2,000 distinct users.

Tools

MongoDB·BigQuery·Python·Apache Airflow·Looker Studio·Google Chat API

← Work