Jaxson Snow

Developed AI-driven personalization systems combining computer vision and LLM-based style profiling for fashion recommendation products. Built multimodal pipelines integrating vision-language models with text-based embedding systems, ensuring consistent cross-modal alignment between image inputs and user style profiles. Implemented trace propagation across vision inference, enrichment layers, and ranking systems using structured event logging and PostgreSQL-backed analytics. Worked across React frontend, API layers, and backend services to unify product behavior across client-server boundaries. Improved recommendation consistency by ~35% and reduced incorrect fallback activations by nearly half through deterministic workflow design and system instrumentation.

Built and stabilized multi-step LLM pipelines involving classification, reasoning, tool-calling, and response generation in production environments with inconsistent failure behavior. Engineered structured workflow orchestration using Node.js, TypeScript, and serverless backend architecture with strict event logging and deterministic execution tracking. Implemented distributed tracing across prompt chains, RAG components, and model routing layers, exposing hidden failure points in fallback and branching logic. Improved incident diagnosis time from 2–3 days to under 1 hour. Designed modular pipeline architecture supporting retrieval-augmented systems, prompt flow engineering, and multi-model orchestration with clear observability boundaries.

Designed and implemented end-to-end observability for production LLM workflows in a fashion AI personalization platform where recommendation outputs were previously non-deterministic and untraceable. Instrumented full-stack pipelines across Node.js, TypeScript, API services, and backend orchestration layers using OpenTelemetry, structured logging, and event-driven tracing. Introduced session-level traceability linking prompts, model versions, retrieval steps, and fallback logic into unified execution graphs. Integrated tools such as Langfuse and PostgreSQL-based event stores for auditability. Reduced debugging time from hours to minutes and cut unexplained output variance by ~60%, enabling safe iteration on ranking and style logic in production.