Work, with context.

• Architected an agentic workflow covenant recommendation system using a 4-node LangGraph DAG with parallel Covenant and Undertaking agents fed by a shared Industry Analyst agent, reducing the business loan assessment cycles from weeks to hours with fully policy-cited, Pydantic-enforced structured outputs.
• Built a self-improving AI system where credit analyst override decisions trigger a LLM judge to extract confidence-scored patterns stored in Amazon Opensearch, retrieved as tool calls during future agent runs, compounding institutional knowledge across 20+ industry verticals and improving recommendation accuracy from 50% to 75%.
• Eliminated hallucinated financial thresholds by designing a two-tier RAG pipeline over Amazon Opensearch with table-aware chunking and re-ranking, enabling agents to cite exact policy sections rather than infer thresholds from priors.

Citing retrieved policy sections makes recommendations easier to review. Retrieval quality and policy freshness still limit correctness.

General rationale for discussing the design. This is not evidence of undocumented project decisions.

• Architected and designed an intelligent release readiness multi-agent system using LangGraph with RAG capabilities, reducing manual validation time by 80% across release management, developer, and platform engineering teams.
• Implemented robust agent orchestration—including state management, error handling, and automated retries—using LangGraph’s graph-based execution to support two monthly releases across 10+ service domains.
• Built a conversational release-readiness agent with advanced prompt engineering, context-aware reasoning, and multi-turn dialogue, leveraging context engineering to synthesize evidence from multiple sources and support queries across four federated business units.

A graph can make dependencies, retries and failure states inspectable. It adds orchestration complexity; checks still need clear ownership.

General rationale for discussing the design. This is not evidence of undocumented project decisions.

• Architected and implemented automated data pipelines to preprocess documents from S3, generate embeddings (OpenAI), and integrate with Amazon Opensearchfor MMR/hybrid search, boosting retrieval accuracy by 30% and reducing processing time by 40%.
• Designed and implemented an end-to-end Generative AI solution achieving 80–90% accuracy in expert-validated insights, enhancing risk analysis and decision-making.
• Developed and refined AI Agents for risk analysis, integrating LangSmith for evaluation and improving generative accuracy and operational efficiency.
• Developed scalable MLOps pipelines using AWS SageMaker, ECR, EKS, and MLflow, enabling comprehensive model versioning, performance tracking, and AI solution reproducibility.

Hybrid retrieval can combine lexical and semantic signals. Improvements need evaluation against representative queries, rather than being inferred from architecture alone.

General rationale for discussing the design. This is not evidence of undocumented project decisions.

• Founded and engineered a GenAI-powered cryptocurrency intelligence platform integrating real-time market data, on-chain indicators, chart images, and sentiment signals for AI-assisted trading research.
• Built multi-modal AI workflows using Google Gemini to generate market summaries, interpret chart patterns, and explain potential price movements from structured and visual inputs
• Designed backend APIs and data pipeline for market data ingestion, AI inference, and analytics delivery, supporting a live production demo.
• Owned full product delivery across architecture, frontend/backend integration, deployment, multi-agent system and user experience.

Combining text, charts and market data offers richer context. Freshness, provenance and uncertainty remain important; output is not a prediction guarantee.

General rationale for discussing the design. This is not evidence of undocumented project decisions.