Projects / Applied AI
A closer look at the problems I work on, the choices behind the solution, and what changed for the people using it.
Case study / Cross-functional AI delivery
A call-center AI assistant has to help while the conversation is still happening. I led the AI and data work and helped align international teams around that real-time need—not just the technical ability to retrieve information.
Less task-processing time per call in the pilot.Reported pilot result for customer-service tasks. This is not a measure of total call duration; the broader call-center rollout is in progress.
The system needed to connect training material, scopes of work, and client records. Conversations with each team showed that retrieval accuracy alone was not enough: a representative needed short, timely, clearly formatted guidance, with the right source behind it.
I worked through the data, user feedback, and real examples with our partners to clarify what mattered: response speed, relevant content, source references, document updates, and real-time transcription and analysis.
Case study / Operational discovery & team delivery
I led a computer-use project to turn operating procedures into a graph-based workflow. The starting documentation described two paths. Testing the process with the people who used it revealed more than ten branches.
Fewer tasks requiring manual handling.Result for the operations workflow covered by this automation—not a company-wide figure. The team also delivered the version ahead of schedule.
I tested the procedure myself, raised questions, and brought Operations and Product into walkthroughs. We documented missing steps, discussed edge cases, and repeated the process as new branches emerged.
I divided the resulting subflows across our AI team. That gave each contributor a defined part of the workflow to implement while keeping the project connected to the way Operations actually worked.
Featured case / Performance & scalability
I led a major update to an existing request-processing pipeline. Each request initially took more than ten minutes. Our target was under two minutes; the update brought processing time below thirty seconds.
As adoption increased, simultaneous requests exposed the limits of a single-threaded design. A full pipeline rewrite could take weeks and introduce migration and versioning risks.
I found an approach that increased concurrency within the existing system. The decision balanced capacity with continuity: improve the experience without a disruptive re-architecture.
The upgraded pipeline supports 20+ concurrent requests, alongside the shorter processing time. It also reduced errors, although no quantified error-rate comparison is published here.
Potential daily throughput increased from hundreds to an estimated 5,000+ requests per day. This describes capacity, not observed traffic. It is separate from the per-request processing time above.
AI agents designed to make progress without hiding uncertainty from the people using them.
Recommendation systems that treat eligibility and operating conditions as first-class parts of relevance.
Machine learning systems designed for shifting data, evolving conditions, and evaluation across time.
Knowledge and communication layers that put the right operational context inside an AI-assisted workflow.
Clarify who acts, what outcome matters, and which constraints cannot be negotiated.
Measure quality, risk, consistency, and failure modes—not only model fluency.
Make outcomes observable so the product, data, and model can improve together.