Principal Data Scientist & AI/ML Engineer
Designing intelligent systems across finance, retail, and real estate
Built a FastAPI + Playwright agent that turns natural spoken commands into structured events, performs conflict checks, and drives the real Google Calendar UI to create or reject meetings.
Created a backend + Next.js front‑end that serves an AI digital twin of myself, grounded in structured facts, LinkedIn profile, summaries, and communication style, with session memory for multi‑turn conversations.
Developed a Flask API around an OpenAI Assistants v2 financial advisor that captures user intent, validates contact details, and writes structured leads directly into Airtable via function calling.
Implemented and evaluated contextual bandit algorithms (LinUCB and baselines) on clinical warfarin data to learn personalized dosing policies and study exploration–exploitation trade‑offs.
Built tabular, linear and deep Q‑learning agents for Pong‑v5, including exploration and learning‑rate schedules, frame preprocessing, and analysis of Atari training curves with TensorBoard.
Implemented a reward model from preference data and used it to train PPO agents with DPO and RLHF variants on the Hopper‑v4 environment, comparing learned vs original returns and analyzing stability.
Implemented vanilla policy gradient with and without baselines across CartPole, InvertedPendulum and HalfCheetah, analyzing variance reduction, learning curves and stability across random seeds.
Implemented a mini GPT model with causal self‑attention, explored cross‑attention and positional encodings, and evaluated how well pretrained Transformers access factual knowledge for question answering over Wikipedia‑style text.
Implemented a bidirectional LSTM encoder–decoder with global Luong attention for character‑level English–Chinese translation, including beam search decoding and careful handling of padding and masking.
A quick overview of my journey across data, AI, and analytics.
Python & JavaScript • Classical ML (regression, classification, clustering, forecasting/LSTM) • NLP (NER, sentiment, transformers/LLMs) • Vision (ViT, CLIP) using PyTorch, TensorFlow, scikit‑learn and Hugging Face.
GCP Vertex AI, BigQuery, Dataflow/Apache Beam, GCS, Pub/Sub • Databricks & MLflow for experiment tracking, model registry and production workflows.
LLM/Agent evaluation (task success, reasoning, safety), A/B testing, drift & quality monitoring, SLO/SLA thinking and rollback strategies.
Node.js, FastAPI, Django and Neo4j for serving APIs, building microservices and modeling complex relationships.
Kubernetes/EKS, Docker, Kustomize, Terraform, GitHub and JFrog for multi‑environment releases and repeatable ML/agent deployments.
LLMs (GPT, Gemini, Llama 3) • RAG with indexing/retrieval/rerank (FAISS, Chroma) • LangChain/LangGraph, prompt chaining, tool use/function calling and agent architectures.
Easy 187 • Medium 212 • Hard 31 • 1,510 problems accepted • 73 daily challenge completions • 2025 global contest ranking around 279k.
For collaboration opportunities, speaking, or just to say hi, feel free to reach out via email or LinkedIn below.