Building LLM & Agentic Systems
for production.

Designed and built major parts of a production Agentic LMS serving 10,000+ learners and 3,300+ daily active users in the U.S., spanning agent orchestration, backend architecture, retrieval, conversational state, streaming, observability, frontend integration, and production infrastructure.

• Built LangGraph-based agent workflows connected to internal tools through MCP, with LiteLLM as the LLM/tool gateway and LangFuse for tracing, prompt management, and evaluation.
• Designed conversational memory and state across PostgreSQL, Redis, Redis Streams, and vector retrieval, including token/message-aware summarization and context compaction for long-running conversations.
• Implemented class-aware hybrid RAG over course materials using vector search, BM25, and RRF reranking; migrated latency-sensitive inter-service calls from REST to gRPC, reducing inter-service latency by ~40%.
• Parallelized independent LangGraph operations and introduced Fast and Advanced modes; Fast mode typically responds in ~2–5 seconds while deeper academic workflows trade additional latency for stronger reasoning and tool execution.
• Built SSE + Redis Streams based real-time and background execution so active agent workflows can continue while users navigate or start another conversation.
• Contributed across React, Next.js, BFF patterns, and reusable components to deliver a seamless real-time teacher/student chat experience.
• Deployed event-driven microservices across GCP and Azure/Kubernetes using Pub/Sub, HPA/KEDA, Docker, GitHub Actions, GitOps, and Argo CD.

Designed and built a production AI-assisted grading platform that has processed 25,000+ student submissions, owning the system from multi-LLM benchmarking and HLD/LLD through implementation, evaluation, deployment, and production optimization.

• Built LLM workflows to parse student submissions, extract structured criteria from complex rubrics and documents, and perform rubric-grounded evaluation.
• Designed adaptive decision flows based on rubric complexity, assignment type, document structure, and available grading context instead of relying on a single fixed prompting pipeline.
• Implemented teacher feedback loops to better align AI-generated grading and feedback with individual evaluation styles, along with human-in-the-loop review for cases requiring validation.
• Benchmarked Claude and Gemini across quality, consistency, latency, and cost using a provider-agnostic architecture.
• Built golden-dataset and RAGAS-based evaluation/regression workflows to make model and prompt changes safer before production rollout.

Designed and owned an end-to-end multi-LLM document extraction platform, converting complex unstructured documents into validated structured JSON for production use across internal products and client systems.

• Built backend services using Python, FastAPI, PostgreSQL, Redis, asynchronous processing, and microservice architectures.
• Designed LLM-based extraction and decision pipelines capable of adapting processing strategies based on document structure, complexity, and extraction requirements.
• Improved backend performance through PostgreSQL query optimization and Redis caching, reducing p95 latency by ~30% across latency-sensitive workflows.
• Applied SOLID principles and reusable design patterns to improve maintainability, provider abstraction, testability, and extensibility.
• Containerized and deployed services with Docker and Kubernetes on Azure/AKS, using Jenkins-based CI/CD and zero-downtime rollouts.

Integrated enterprise SaaS systems into a production GenAI assistant, supporting workflows across an ecosystem of 50,000+ users.

• Integrated NetSuite, Shopify, UPS, and Avalara into a GenAI assistant using intent classification, RAG, and backend orchestration.
• Built workflows around product information, order status, availability, shipping, taxation, and related e-commerce operations.
• Optimized high-latency NetSuite operations using asynchronous execution and Redis caching, reducing response time from ~8s to ~1.2s — approximately a 7× improvement.