Selected projects

Career guidance inside an enterprise has to combine employee context with internal knowledge while keeping AI behavior constrained enough for employee-facing use.

Designed the AI architecture for a conversational career coach using Azure AI Foundry and Azure OpenAI, with enterprise RAG, orchestration, guardrails, and content-safety controls around personalized career guidance.

A production-oriented architecture for career guidance, skill-gap analysis, learning recommendations, and promotion-readiness support within a secure enterprise environment.

[ Azure AI Foundry // Azure OpenAI // Enterprise RAG // Guardrails // Content Safety ]

Federal criminal-defense drafting depends on hierarchical legal sources, accurate retrieval, and clear linkage between arguments and the material supporting them.

Built the platform from scratch, including specialist agents across different legal corpora, triage and routing, semantic retrieval with Qdrant, keyword retrieval with Elasticsearch, and source-linked drafting. A separate evaluation node scored drafts against a custom rubric and returned structured feedback for another iteration when required.

Drafting that previously took attorneys many hours could be produced in minutes while preserving the evidence trail behind the generated arguments.

[ Multi-Agent Orchestration // Qdrant // Elasticsearch // Rubric Evaluation // LangSmith // AWS ]

Generating marketing assets is easy; identifying the right personas and angles for a product before generation is the harder part.

A staged agentic pipeline that moved from product understanding to persona selection, angle selection, and asset generation, with each stage passing structured context into the next.

Asset generation became driven by explicit audience and positioning decisions rather than a single generic generation step.

[ Multi-Agent Orchestration // Persona Selection // Positioning // Structured Handoffs // LLMs ]

Unstructured documents needed to be converted into structured knowledge, but the entities and relationships to extract changed depending on the ontology defined for each use case.

An ontology-driven extraction pipeline that identified schema-aligned entities and relationships from documents and mapped them into a knowledge graph.

The same extraction architecture could support different domain schemas without forcing every document into one fixed representation.

[ Knowledge Graphs // Ontology-Driven Extraction // Neo4j // LangChain // LlamaIndex ]

Investors needed useful summaries and signals about individual crypto projects from very large volumes of noisy Telegram and social-media data.

Designed NLP pipelines that extracted key entities, sentiment, topics, relationships, and activity patterns, then summarized the information around each crypto project for investment research.

Large conversational datasets were turned into project-level intelligence, while the architecture balanced model quality against the latency and throughput required at scale.

[ NLP // Fine-Tuned NER // Sentiment Analysis // Summarization // Clustering // Graph Analysis ]