•Architected the multiagent generation pipeline behind B12s AI website draft flow, including LLM
orchestration, model routing across OpenAI and Anthropic, schema-validated structured output, and
retry/fallback handling; reduced median time-to-first-draft from ~6 minutes to under 60 seconds.
•Designed the retrieval-grounding layer using PostgreSQL and pgvector, anchoring generated marketing
content to scraped business data and curated industry corpora; reduced hallucinated service claims identified
during editorial QA by ~65%.
•Built the streaming authoring experience in React and TypeScript using server-sent events and optimistic
updates, allowing users to see website sections generated and assembled in real time.
•Established LLM evaluation using golden datasets, LLM-as-judge scoring, and CI regression gates to validate
prompt and model changes before release.
•Reduced inference cost by ~40% per generated site through smaller-model routing, semantic response
caching, and batched embedding writes, without measurable degradation in evaluation scores.
•Mentored three engineers and established development standards for prompt versioning, fallback behavior,
structured outputs, and safe LLM integration.
•Implemented the browser-based EON-XR lesson authoring tool using React, Three.js, and WebXR, allowing
instructors to add hotspots, quizzes, and narration to interactive 3D models without writing code.
•Built the glTF/FBX conversion and LOD pipeline as a C#/.NET service, reducing 3D model load times from
~12 seconds to under 3 seconds on mid-range mobile devices.•Integrated speech-to-text transcription and automatic quiz generation into the authoring workflow, reducing
narrated-lesson production time from several hours to ~20 minutes.
•Delivered SCORM and LTI integrations that enabled EON-XR lessons to be published through university LMS
environments.
•Contributed to Unity/C#, WebGL, and cross-platform delivery, eventually taking ownership of the asset
pipeline modle
•Owned end-to-end delivery of 3D product configurators on the Jikannai platform, from NestJS/TypeScript
APIs to React customer portals, enabling customers and sales teams to configure, quote, and order complex
outdoor products.
•Implemented product constraint engines for additions such as gates, ladders, and kits, enforcing placement
and compatibility rules before configurations reached quote generation.
•Built multi-tenant organization and authorization models with role-aware APIs, supporting manufacturer
administrators, dealers, and end customers with tenant-scoped catalogs and pricing.
•Replaced synchronous quote and export operations with Redis-backed job processing and webhook
callbacks, eliminating request timeouts that were a major source of support issues for large orders.
•Reduced configurator p95 interaction latency from ~4 seconds to under 1 second through query
optimization, cursor-based pagination, and caching of geometry and BOM lookups.
•Increased backend test coverage from ~40% to ~85%, improving confidence in changes to core platform
functionality.