Work

Context

Internal manufacturing teams needed a reliable AI system for a complex, domain-specific operational workflow involving multiple roles and existing software.

Responsibility and collaborators

I work across problem framing, architecture, implementation, evaluation, deployment, and iteration with manufacturing stakeholders, domain experts, architects, data scientists, data engineers, and software engineers. My official title remains Data Scientist; within the initiative, I provide technical leadership across the system’s development.

System approach

The solution is a complex multi-agent application operating in a manufacturing environment. My work includes coordinating agent responsibilities, application logic, structured state, evaluation, human oversight, cloud deployment, and enterprise integration.

Engineering decision

A central design decision was to treat language models as components inside controlled software rather than as autonomous systems. Reliability comes from clear boundaries, validation, review, and explicit failure handling.

Evaluation

Evaluation focuses on task completion, consistency, constraint adherence, failure modes, and whether domain experts can use the system effectively. Findings from user review are incorporated into subsequent iterations.

Stage

Active internal engineering work with manufacturing stakeholders. Deployment and integration continue.

How these systems are grounded, constrained, and evaluated →

Scope of contribution

• Technical architecture
• Multi-agent application engineering
• Evaluation and failure analysis
• Human oversight
• Cloud deployment
• Cross-functional delivery

Independent mental-health professionals often manage scheduling, billing, receipts, reminders, and client information across disconnected tools. OneClick is the administrative platform I co-founded to collapse that work. It is not clinical documentation software.

I own product decisions, architecture, engineering, security, deployment, and operations, alongside CAE work. The product was co-designed with practicing clinicians — seven rounds of sessions with 38 mental-health professionals.

Public launch is planned for October 2026. The full product case, including interface screenshots, lives on the OneClick page.

OneClick product case →

Before the current agentic work, operational teams needed machine-learning systems that could forecast demand and support maintenance decisions from complex time-series and equipment data.

I designed, trained, validated, and helped put those models into production with domain experts and engineering counterparts: deep-learning predictive maintenance, demand forecasting, large-scale data processing, scheduled inference, and monitoring. Related earlier CAE work included speech and language-model systems for noisy environments and technical documents.

Production ML notes →