02 / WorkSelected case studies

Applied AI, in operating context.

A small set of case studies organized around problems, decisions and lessons—not a chronology. Details stay intentionally bounded to protect confidential work.

LLM · NLP · RAG

LLM systems for e-commerce operations

Production-oriented NLP and LLM systems integrated into operational workflows.

Problem: Operational teams needed faster access to fragmented context and more consistent support for recurring decisions.
Context: The work sat inside real workflows, where retrieval quality, response latency, access boundaries and failure handling mattered as much as the model itself.
My role: Shaped applied AI use cases, connected model behavior to product requirements and worked across RAG, agents, multimodal prototypes and production integration.
Approach: Started from the decision being supported, designed retrieval and tool interfaces around that decision, and added explicit boundaries where model output should remain advisory.
Result: Built AI and NLP flows that reduced response time in selected operational workflows from minutes to seconds.
Lessons: Useful LLM products depend on context quality, observable failure modes and a clear human decision point—not only a stronger base model.

Agents · LLM · Tools

Multi-agent inventory platform

A coordinated system for merchant inventory operations with specialist agents and human confirmation.

Problem: Inventory questions crossed several operational domains and systems, making a single undifferentiated assistant difficult to control and evaluate.
Context: Supply, transit and export workflows each required their own context and tools, while critical operations had to remain under explicit user control.
My role: Designed the agent topology, coordination flow and confirmation boundary for operational actions.
Approach: Routed requests through a coordinator, delegated bounded tasks to specialist agents, retrieved context from internal systems and returned a proposed action for review.
Result: Produced a coherent operational proposal while keeping critical execution behind an explicit confirmation step.
Lessons: Specialization helps when tool scopes and responsibilities are real. Coordination becomes reliable only when agents share a clear contract and escalation path.

User request

↓

Coordinator agent

↓

Supply agent

Transit agent

Export agent

↓

Internal systems & tools

↓

Proposed action → explicit confirmation

Strategy · MLOps · Architecture

AI-first transformation architecture

A maturity model that connects process digitization to data, ML platforms, LLMs and agents.

Problem: Teams often approach AI as an isolated feature while the processes, decision history and infrastructure it depends on remain implicit.
Context: Sustainable AI adoption requires both a technical platform and an organizational sequence for turning work into learnable, governable systems.
My role: Worked on AI transformation strategy and the architecture that joins business rules, datasets, MLOps and applied AI products.
Approach: Used a maturity chain to identify missing foundations and sequence investments before introducing model-driven automation.
Result: A shared framework for discussing where an organization is ready for AI and which prerequisite should be strengthened next.
Lessons: An agent cannot compensate for missing process data or unclear business rules. Transformation starts by making decisions legible.

Paper
Digitalization
Decision history
Business rules
Dataset
ML
MLOps
Infrastructure
LLM
Agents