Built Kubricks, a React/TypeScript + Flask platform for evaluating multimodal AI video tasks with Celery/Redis async pipelines for video ingestion, frame extraction, and Whisper transcription.

Constructed multimodal prompts combining transcripts, extracted frames, and task instructions, invoking Gemini via LangChain with retry/batching logic and structured JSON output parsing.

Developed automated evaluation pipeline comparing model responses against gold annotations, detecting hallucinations and temporal errors with PostgreSQL analytics dashboards.

Implemented token usage and latency tracing with replay functionality for rerunning tasks under modified prompts or model configurations.

Built and shipped core React Native features (guest lists, RSVPs, creator discovery) used by 50k+ users, improving mobile performance by 30% through optimized rendering.

Developed scalable Node.js backend APIs for high‑volume event workflows, reducing user drop‑off during event flow by 25% with improved validation and error handling.

Optimized database queries and implemented Redis caching for event search, reducing API response times from 800ms to 120ms (85% improvement) and supporting 500+ concurrent users.

Deployed and scaled backend services on AWS (EC2, RDS), enabling reliable handling of high-concurrency traffic and improving system uptime.

Built clinical AI nutrition platform serving 200+ patients with subgraph DAG + RAG pipeline (LangChain, QDrant), improving dietary recommendation accuracy by 22%.

Collaborated with Food Network Chef James Briscione and White House Fellow Prof. Lav Varshney to integrate culinary expertise into AI recommendation engine.

Deployed HIPAA-compliant Next.js platform with GPT-4o integration, implementing encryption, audit logging, and secure PHI handling, passing UCSD Health IT security review.

Built agentic workflow with LangGraph for multi-step reasoning (medical history, allergies, meal generation), reducing manual dietician review time by 65%.

Built TypeScript backend (Bun + Hono + Postgres) with unified data layer, multi-tenant model with gateway-enforced tenantId, and source-native JSONB storage.

Designed provider abstraction layer with incremental sync pipelines using async generators, centralized upsert system (upsertEntity), and scheduler with concurrency control.

Implemented source-of-truth cache with entities table (tenant, source, entity_type, external_id), checksum-based diffing, and webhook ingestion with provider validation.

Built tool-based agent system with per-session isolation, WebSocket streaming for real-time responses, and context compaction for large token windows.

Implemented core primitives: Jobs (long-running agent tasks), Alerts/Triggers (event-driven notifications), Pins (persistent visualizations), and Workspaces combining all three.

Enabled 4+ enterprise clients processing $2M+ GMV, helped raise $1.0M seed round, and created extensible system for cross-tool automation.

Built full-stack ETL dashboard processing 120k+ clinical trial reports, reducing document analysis time from 4 hours to 45 minutes per batch (80% reduction) with automated extraction.

Implemented semantic search using OpenAI embeddings and FAISS vector database, improving R&D team productivity by 3x for finding relevant trials and adverse event reports.

Built FastAPI backend with async processing for large-scale ingestion, implementing multi-threaded PDF parsing and OCR (Tesseract) for scanned documents.

Upgraded PostgreSQL + AWS S3 stack with database optimization and indexing, increasing throughput 4× (500 to 2,000 documents/hour) for R&D pipeline.

Modeled ion–electron Coulomb drag coupling to predict open‑circuit voltage and short‑circuit current in Si nanochannels.

Designed CNT‑integrated channel architectures targeting higher momentum transfer and energy harvesting efficiency.

Built subgraph DAG + RAG pipeline (LangChain, QDrant vector DB), improving dietary recommendation accuracy by 22%.

Collaborated with Food Network Chef James Briscione & White House Fellow Lav Varshney on clinical AI nutrition engine.

Built real‑time React + Flask platform for multi‑user LLM interaction research (40+ study subjects).

Ensured zero‑loss message reliability via WebSockets + analytics logging for human–AI interaction experiments.

Integrated RL agent into MPC framework for autonomous EV routing in SUMO → improved control accuracy 25%.

Achieved 18% lower traffic delay in open‑world simulations using ROS, C++, Python.

Analyzed experimental data and developed new concrete mixture by reducing clinker in cement.

Built Python + Random Forest models to predict composition impact on durability, reducing carbon emissions by 10%.

Held office hours and provided one-on-one support for students in computer science courses.

Graded assignments and provided detailed feedback to help students improve their programming skills.

Present technical presentations on hardware design, quantum computing, and machine learning applications in engineering.

Collaborate on circuitry and embedded systems projects with fellow IEEE members and technical communities.

Conducted market analysis and developed a go-to-market strategy for a new product line in the pipefitting industry.

Used NLP techniques to simplify complex insurance documents, making them more understandable for customers.

Developed and implemented a secure UART communication protocol for an embedded medical device.

Contributed to securing x86 architecture firmware by implementing memory-safe coding practices in Rust.

Built and styled responsive UI components for a cross-platform mobile app using React Native.

Translated Figma mockups into functional application screens and integrated them with a Python Flask backend.