Companion to Work
AI Engineering
Engineering judgment behind reliable enterprise AI: multi-agent architecture, grounding, evaluation, human oversight, and integration with existing software. Read the manufacturing case →
Grounding and enterprise knowledge
Useful internal systems need access to proprietary knowledge that is not in the model weights. Grounding is treated as an engineering problem: what to index, how to retrieve it for a given task, and how to show the user what the system used.
Multi-agent architecture and constrained execution
Language models are components inside controlled software, not autonomous systems. Work I do in this area includes coordinating agent responsibilities, keeping structured state, and returning outputs that can be reviewed rather than treating every step as free-form generation.
Reliability comes from clear boundaries, validation, human oversight, and explicit failure handling. Fallback paths exist when the system should not proceed.
Multi-agent architectureCoordinated software with defined responsibilities.
GroundingEnterprise knowledge outside the model weights.
Structured stateApplication logic that can be inspected and constrained.
Human oversightReview, approval, and fallback when needed.
EvaluationTask completion, consistency, and failure analysis.
Enterprise integrationCloud deployment alongside existing software.
Evaluation, failure analysis, and iteration
Task-level checks ask whether the system completed the assigned work, followed constraints, handled failures, and produced something a domain expert can use. Failures are inspected rather than averaged away.
Regression checks and review with users close the loop. That is closer to engineering validation than to a one-off demo.
Deployment and integration
Current agentic work includes cloud deployment and integration with existing enterprise systems. Some integration work remains in progress; this page does not claim a completed production rollout of every workflow.
Earlier production machine learning at CAE included forecasting, predictive maintenance, scheduled inference, and monitoring inside operational workflows. Those systems are a different stage from the current agentic work.
Selected technologies
Tied to verified work, not used everywhere: Python, PyTorch, Azure, Databricks, PySpark, FastAPI, Docker, SQL. OneClick uses Django, PostgreSQL, Azure containers, Stripe, and AWS SES. React appears in product engineering where a web interface is required.
Human review
Sensitive workflows keep explicit review, approval, and fallback paths. Privacy and access constraints are treated as design limits, not afterthoughts.