Leopoldo – Python, LLM, AI agent orchestration

A production AI credit-scoring platform that extends loans to smallholder farmers across Uganda, using alternative data to assess creditworthiness when conventional credit history is unavailable. It scores applicants, automates lending decisions, and serves a fast-growing base of rural borrowers.

• Designed and deployed the Azure-based MLOps pipeline, automating model training, validation, and deployment for the production credit-scoring models.
• Built the Medallion-architecture DataOps layer (bronze/silver/gold) and the feature-engineering pipeline feeding the models, replacing ad-hoc data handling with reproducible, auditable datasets.
• Defined automation and model-governance standards with business stakeholders to keep lending decisions explainable and compliant.
• Mentored the local engineering team to own and operate the platform independently after handover.
• Laid the DataOps and feature foundation before automating training, so every model run traces back to versioned, validated data — essential for a regulated lending product.

This role required spearheading the transformation and marketing of AI products, driving their sales potential.

• Enhanced advanced AI and digital human research into commercial offerings ready for enterprise deployment.
• Directed productization of both platforms.
• Negotiated multimillion-euro partnerships with Sony, Ubisoft, Electronic Arts, and Katapult Group.
• Built cross-functional AI and design teams.
• Defined go-to-market architecture for human-centric AI solutions.

An NLP-driven conversation engine that powers responses and animation for virtual humans. It pairs tag-based matching with fuzzy database search and multi-armed-bandit decision logic to select the best response in real time — delivering human-centric conversational AI without the cost or training overhead of deep-learning models. Taken from research prototype to enterprise-deployable product.

• Directed the architecture and productization of the product from prototype to an enterprise-deployable conversation engine.
• Designed it as a tag-driven matching system combining fuzzy search with multi-armed-bandit decision logic to pick the best response in real time.
• Defined the technical go-to-market architecture — deployment, integration, and data-protection models for enterprise buyers.
• Owned the technical side of due diligence in multimillion-euro partnership talks with Sony, Ubisoft, Electronic Arts, and Katapult Group.
• Chose tag-matching and bandits over an LLM/deep-learning pipeline to get real-time recommendation quality with no training cost or inference latency.

A system for generating and managing character animation at scale — the production counterpart to DELPHIC for virtual-human products. It uses tag-based matching to pair animation assets to context in real time, letting teams drive large animation libraries without manual, per-asset selection.

• Directed the architecture and productization of the product from prototype to an enterprise-deployable product.
• Built it as a tag-driven matching engine pairing tag selection with fuzzy search to surface the right animation asset in real time.
• Integrated the platform with DELPHIC so conversation and animation are selected together for coherent virtual-human behavior.
• Owned engine-level technical direction across the animation toolchain.

A multi-tenant SaaS RAG chatbot platform that turns an organization's contracts and internal documents into a trustworthy, queryable knowledge base. It ingests and embeds each tenant's documents, automatically detects contradictions and conflicts across them, and surfaces a reconciled knowledge layer that users can query through a chatbot. Built for teams that need their document corpus to stay consistent and reliable as it grows.

• Architected and built the full platform in Rust (Axum) with PostgreSQL and pgvector, including tenant isolation, document ingestion, and embedding pipelines.
• Designed a multi-level agentic pipeline that catches contradictions in contracts and legal documents, sanitizes content, and lets users resolve conflicts before they reach the chatbot.
• Integrated the knowledge-conflict detection into a Harvey.ai workflow, restructuring the agent flow to cut LLM token cost.
• Built the entire Azure orchestration for multi-tenant, multi-project deployment.
• Chose Rust and pgvector over a Python and managed-vector-DB stack to keep per-tenant latency and memory predictable at scale without external vector-database costs.

An AI-powered health and behavior-change platform that reached tens of thousands of users with ML-driven coaching. It combines a real-time decision engine, content classification, and recommender systems to guide each user to their next best step — surfacing the right content, intervention, or distraction at the right moment across a continuously growing personal profile.

• Architected a distributed, event-driven cloud platform on Azure spanning mobile clients, backend services, and ML-driven coaching logic.
• Designed SYBIL, a real-time decision engine built on proprietary swarm logic — a pathfinding-style system that recommends each user's next best step with no model training required.
• Built "Shadow Habtonomics," an NLP content-classification system that engineered features from unstructured content (type, tone, subject) and matched each piece to a user's profile and prior responses.
• Developed three complementary recommender systems: the most therapeutic path, real-time detection of when a user needed distraction (surfacing mini-games), and deliberate exposure to content they would normally avoid.
• Scaled the team from 10 to 120 across 40+ annual hires and led the Agile transformation while holding delivery quality stable.
• Chose a swarm/pathfinding decision engine over a trained recommender so guidance worked from day one on sparse data and stayed explainable as each user's profile grew.

A healthcare operations software — ERP, electronic patient-record, and workflow-management systems — for clinical care providers such as RugHuis and PHI. Beyond the operational platform, it applied machine learning to admissions: a classification model predicted which applicants were most likely to complete their treatment, helping clinics direct scarce capacity to where it would do the most good in a setting where demand vastly outstripped supply.

• Architected and launched ERP, electronic patient record, and workflow-management systems for healthcare clients RugHuis and PHI, leading a 28-person multidisciplinary team.
• Built a classification model predicting each applicant's likelihood of completing their treatment trajectory, engineering features from demographic, economic, and social markers alongside intake-form data on each applicant's specific problem type.
• Used the model to prioritize admissions where demand ran roughly 40x available capacity, directing limited spots to the applicants most likely to succeed.
• Developed the 3D avatar and AI prototypes that became the technological core of Dephion Group.