Mark – Python, React, Next.js

Speech-to-speech conversational learning tool (LTI) embedded in an LMS, built for AI-driven educational interactions. The platform supports multi-agent Dialogue flows, real-time voice and text sessions over WebRTC, document-aware tutoring with in-session RAG, and AI-generated Key Moments for post-session review. Deployed on AWS and GCP with full accessibility (ARIA/WCAG) compliance.

• Architected multi-agent AI systems using LangChain, LangGraph, and OpenAI — enabling real-time user-to-agent voice and text interactions via WebRTC with supervisor/worker agent patterns
• Replaced Whisper-based transcription with a custom decoupled layer that preserved accuracy when agents were interrupted; added timestamp-based message sorting across LTI and LMS sessions
• Implemented in-memory vector stores with OpenAI Embeddings and pgvector for near-zero-latency real-time RAG over session-bound, multi-source documents
• Built the Dialogue system — a multi-step, context-carrying multi-agent framework with shared conversation history, a sequence builder for assignment types, and AI-generated Key Moments for post-session review
• Instrumented LLM pipelines with Langfuse for end-to-end tracing, latency profiling, and token usage optimization across multi-agent production flows
• Deployed AI-powered services end-to-end (Python/FastAPI, Go/GORM, Next.js) on AWS and GCP, from architecture through production monitoring

Personal finance management app for tracking expenses, income, and financial goals, with interactive dashboards and data visualizations. Full-stack open source project with production-grade security architecture.

• Built a full-stack finance tracker with a Next.js 15 / React 19 frontend and NestJS REST API backend backed by PostgreSQL
• Implemented interactive charts and expense visualizations using Chart.js
• Set up JWT authentication with CSRF double-submit cookie protection, Helmet headers, rate throttling, and input sanitization
• Containerized the full application with Docker Compose for reproducible local and deployment environments
• Maintained 215+ commits with branch-based workflow, PR reviews, and documented architecture

Energy management software for a transformer monitoring system (SMT) at one of the world's largest power plants. The platform tracked real-time sensor data from high-voltage transformers, with strict zero-downtime and regulated testing requirements.

• Refactored an IoT anomaly detection system with AI-driven predictive analysis, improving fault detection accuracy for high-voltage transformer monitoring
• Built serverless and microservices architectures on AWS (Lambda, DynamoDB, S3, EventBridge) processing real-time sensor data streams via Kafka
• Developed real-time data visualization dashboards for the Energy Management sector using Next.js and D3.js
• Optimized PostgreSQL queries and Redis caching layer, reducing system latency by 13%

Internal tooling team focused on Node.js ecosystem modernization and CI/CD infrastructure. Worked on migrating JavaScript codebases to TypeScript and improving developer tooling used by tens of thousands of engineers daily.

• Wrote unit and integration tests for an internal CI/CD system on the Node.js tooling team, achieving 95% test coverage
• Implemented a Redis caching layer for session tokens in a legacy microservice, reducing average auth latency by 100ms
• Migrated legacy JavaScript utilities to TypeScript and improved CLI tooling serving 20K+ daily users