• LLM platforms & agents

  • Fine-tune and serve open-source LLMs (Axolotl + QLoRA) behind vLLM for enterprise agent use cases.
  • Design guardrailed RAG / GraphRAG systems with content safety (OpenAI Moderation, NeMo Guardrails) and retrieval metrics baked in.

• ML systems & explainability

  • Ship model-agnostic attribution and PDP pipelines using DuckDB + Arrow, aligned with LightGBM + SHAP outputs for production dashboards.

• Data & infra

  • Build and harden FastAPI services, K8s reconciliation loops, and data pipelines in SparkSQL, PostgreSQL, and warehouses.

I care about robustness, observability, and measurable impact – not just getting a Jupyter notebook to “work.”

• Custom LLM fine-tune & serve platform

  • Architected a fault-tolerant fine-tune-and-serve platform (Axolotl + QLoRA + vLLM) for multiple open-source foundation models.
  • Automated checkpoint detection & recovery, cutting manual setup / monitoring by ~80% and enabling a Fortune 50 telecom to ship a security metadata classifier on time.

• Model-agnostic feature attribution & PDP

  • Built a single-pass explanation pipeline using model predictions + univariate summaries with DuckDB pivoting over Arrow data.
  • Unified global + pairwise explanations across LightGBM and SHAP, and reduced per-feature PDP compute 17× while keeping curve fidelity (ρ≈0.90).

• Model reconciliation controller

  • Re-architected a fragmented Flask process into a FastAPI proxy with an idempotent reconcile loop using DeepDiff + K8s helpers.
  • Consolidated four manual steps into one API and cut manual recovery effort by 90%, speeding pod rollouts 3× and eliminating double-launch incidents.

• Graph-structured document summarization

  • Implemented RAG + GraphRAG over Neo4j with embedding clustering, token-based splitting, BM25, MMR, and FlashRank-based re-ranking.
  • Integrated OpenAI Moderation + NeMo Guardrails and improved retrieval nDCG@k by 25%, shipping a modular retrieval suite into production.

• PPO post-training for Llama text-to-SQL (3B)

  • Engineered a PPO pipeline that boosted F1-SQL from 16% → 84% using ~1k human feedback samples and ≈$11 of H100 compute, reaching task-bounded parity with an OpenAI o3-series model.

• Prior ML work

  • SATD detection and refactoring recommendation (capstone @ RIT).
  • Histopathology carcinoma classification using multi-level spatial fusion (CCIS book series, FTNCT 2019).
  • Long-form writing on Medium (Towards Data Science, The Startup, etc.) on topics from customer segmentation to the last 40 years of gaming.

Languages
Python · SQL · Cypher

ML / LLM
PyTorch · TensorFlow · Keras · scikit-learn · LightGBM · SHAP · ONNX
Axolotl · QLoRA · vLLM · LangChain · LlamaIndex · NeMo Microservices · NeMo Guardrails
OpenAI · Vertex AI

Data & storage
DuckDB · PostgreSQL · MongoDB · Neo4j · Chroma · PySpark
AWS · GCP

Backend / infra
FastAPI · Flask · Kubernetes · Docker · GitHub Actions

Some older but representative public work:

• Gaming-Industry-Analysis
  Data analysis of a 40-year gaming dataset – genre/platform trends, sales patterns, and publisher contributions, plus a companion blog post tying it all together.

• Prediction-of-Customer-Churn
  ANN-based churn prediction for banking customers, with extensive visualization (ROC, confusion matrix, pie charts, KDE, counter plots).

• Disaster-Response-Pipeline-Web-App
  End-to-end ETL + NLP + ML pipeline that powers a web app for classifying disaster-related messages into multiple categories.

• Recommendation-of-Refactoring-Techniques-to-address-Self-Admitted-Technical-Debt
  SATD detection and refactoring recommendation – the public side of your RIT capstone work.

• The Essential Guide to Effectively Summarizing Massive Documents, Part 1
• Advancing the Power of Retrievers in RAG Frameworks
• Customer Segmentation, Identifying the Profit Among the Loose Ends.
• The Last 40 Years of Gaming Industry, Unlocked.

_{This section is auto-updated from my Medium RSS feed.}

• Authored the core problem statement & evaluation metrics for the UC Berkeley AI Summit 2023 – Data Science Hackathon.

• Represented Aible at Ai4 2023, Google Next 2024, and AWS Summit 2024, running technical demos and stakeholder-facing discussions.

• Enjoy long-form writing on data, ML, and games whenever I can find the time.

_{⚡ Fun fact: I will absolutely over-analyze both fragrance notes and video-game industry trends.}