AI & ML Engineer · Data Engineer · LLM Research · Dayton, Ohio, USA

Email · GitHub · LinkedIn · Hugging Face

I build production-grade intelligent systems end-to-end — multi-agent LLM platforms, RAG evaluation pipelines, distributed CDC architectures, GPU inference engines, and full-stack applications — and research low-resource language modeling and LLM fairness on the side. Currently open to AI/ML Engineer and Data Engineer roles.

AI Engineering

• production-rag-eval-platform — RAG evaluation platform benchmarking 3 retrievers against 3 LLM providers; 92% faithfulness and 100% retrieval hit rate, with GitHub Actions gates blocking prompt/retrieval regressions.
• FinSight-AI — 4-agent async financial research system across US, NSE, and BSE markets, combining FinBERT sentiment, LLM-as-judge scoring, and Markowitz portfolio optimization.
• Vocalytics — Voice-to-SQL BI copilot over 3.4M orders / 30M+ line items; guarded SQL generation with dynamic allowlists and a 29-test eval suite.
• tok-adapt — Python/CLI library for tokenizer and embedding-layer adaptation, plus an end-to-end CPT → SFT → DPO pipeline for cross-lingual LLM fine-tuning.

Data Engineering

• cdc-pipeline — Exactly-once CDC pipeline sustaining 50,000 writes/min: PostgreSQL WAL → Debezium → Kafka → Flink → Apache Iceberg, chaos-tested against TaskManager kills and network partitions.
• data-lakehouse — Delta Lake compaction engine cutting per-partition file count 95% (87 → 4 files) in 14.5s with zero reader downtime.
• data-quality-ai — Real-time data quality and LLM incident-triage platform detecting 6 failure modes; HIGH-severity incidents flagged in under 42ms.

ML Engineering

• tensorrt-vision-server — ResNet-50 compiled to an FP16 TensorRT 11 engine on CUDA 12.4; 3.55× throughput (97 → 347 FPS), 4.3× P99 latency cut, zero accuracy loss.
• feature-store-skew — Feast/PySpark/Redis feature store enforcing train-serve parity, with SHA-256 hash comparison catching skew at inference time.
• churn-ml — XGBoost churn pipeline with 50-trial Optuna tuning, an MLflow Registry promotion gate, and Evidently drift monitoring.

Full-Stack

• flowforge-automation-engine — Visual workflow automation platform: React Flow editor, Spring Boot DAG executor, JWT auth, and OAuth credentials encrypted at rest (Jasypt AES-256).
• flash-sale-reservation-platform — High-throughput flash-sale reservations with Redisson Redlock inventory holds, Kafka-backed checkout, and live seat availability over WebSocket.
• quire-workspace — Real-time collaborative document editor with live multi-cursor editing via Yjs CRDTs, on a Spring Boot WebSocket hub.

Independent research in low-resource language modeling and LLM fairness, alongside my engineering work.

• How Fast Can Reward Models Score? — a systems study of C++ and PyTorch inference runtimes for RLHF reward scoring, finding the CPU speedup traces to ONNX Runtime's graph execution rather than C++ itself. arXiv:2607.19712 · code
• Evaluating Dedicated Monolingual and Joint Multilingual Causal Models for Dravidian Languages — 5 GPT-2-architecture models for Telugu, Tamil, Kannada, and Malayalam, benchmarked against XLM-R, mBERT, and mGPT on tokenizer efficiency and downstream transfer. arXiv:2608.07727 · code
• Release-Level Fairness Drift in Large Language Models — a fairness-regression audit pipeline across 12 model versions from 5 providers, 6 model families, and 7 bias benchmarks (~200K examples). Under double-blind review. code

AI/ML — Python, PyTorch, TensorRT, CUDA, ONNX, Hugging Face Transformers, scikit-learn, XGBoost, Optuna LLM/NLP — Claude, OpenAI, Ollama, ChromaDB, Sentence Transformers, SentencePiece, Whisper, FinBERT Data Engineering — Kafka, Spark, Flink, Delta Lake, Iceberg, Debezium, PostgreSQL, Redis, DuckDB Full-Stack — Java, Spring Boot, React, TypeScript, Yjs, WebSocket, JWT MLOps/Infra — MLflow, FastAPI, Docker, GitHub Actions, Feast, Evidently, Prometheus

Open to collaborations, open-source contributions, and mentorship in AI/ML, data engineering, and NLP research — reach out on LinkedIn or email me.