Senior Software Engineer / AI Engineer at Together AI

Built a support ticket platform using a multi-step AI agent system with LangGraph and Claude Sonnet to automate ticket triage and report generation. Modeled the workflow as a stateful graph (plan → call tools → evaluate → escalate), with checkpointing to resume after failures or human approval. Integrated the agent with external systems via typed tool schemas (Zendesk/Jira ticket data, internal REST APIs for account/billing data) and Slack for human escalation, enabling model-validated actions. Owned a full-stack surface exposing LangGraph runs over streaming endpoints so operators could watch and interrupt automation mid-run. Reduced cost and latency by routing simple subtasks to a smaller model while reserving Sonnet for planning/reasoning steps. Improved end-to-end task success rate using a retry/evaluation layer and routed to human review when confidence was low.

Senior Software Engineer at Hippocratic AI

Designed a RAG pipeline using LangChain and Pinecone to ground a patient-facing voice AI agent in clinician reference materials, reducing unsupported or hallucinated claims by grounding generated statements in retrieved, citation-backed chunks. Implemented a document ingestion pipeline with metadata tagging (specialty, source, effective date) to support filtered retrieval by call type. Built hybrid retrieval (dense embeddings via Pinecone plus keyword/BM25 reranking) and relevance threshold gating so that when no sufficiently confident match was found, the agent escalated to a human nurse instead of answering from low-confidence or empty retrieval. Added a dedicated verification/safety sub-model that cross-checked generated responses against retrieved chunks before the voice agent spoke. Owned retrieval evaluation and regression testing, including building a golden dataset of clinician-reviewed Q&A pairs by specialty and automating retrieval quality checks (recall@k, faithfulness) in C

Software Engineer at Microsoft

Built meeting scheduling and calendars sync features for the Microsoft Teams web client by implementing backend services in Node.js/TypeScript to synchronize availability across Outlook, Exchange, and Teams, scaling calendar sync via internal Graph APIs. Improved the chat message search backend by replacing linear database scans with an indexed search service, reducing average search latency for users with large front-end data and preventing redundant API calls while typing. Migrated legacy jQuery-based admin settings to a React/TypeScript panel with strict typing to reduce null-reference production bugs and improve maintainability for other teams. Built a real-time notifications service using Node.js and WebSockets to push meeting reminders and @mentions to the Teams client, handling connection/backoff logic for unstable networks and instrumenting delivery latency telemetry. Partnered with design/accessibility teams to support WCA G2.1 AA compliance with appropriate ARIA roles, keyboa