I don't just train models; I engineer resilient, deterministic AI systems.
My passion lies in taking fragile LLM prototypes and hardening them into scalable, agentic microservices. Whether it's architecting graph-based Self-RAG pipelines with zero-cost NLI guardrails, publishing open-source data tooling, or optimizing low-latency vector databases, I operate at the intersection of high-performance engineering and applied cognitive intelligence.
- 🎓 MS in Computer Science from Kent State University (GPA: 3.97 / 4.0) | BS in CS (GPA: 3.7 / 4.0).
- 📜 Researcher with peer-reviewed publications across IEEE CAI 2026, MDPI AI (2025), and SKASE (2024).
- 📦 Open-Source Tooling: Creator & Maintainer of
wranglefyandtprdbreaderon PyPI (80% dataset parsing speedup). - 🛡️ Certifications: Anthropic Model Context Protocol (MCP) Specialized Engineer, AWS Machine Learning - Specialty, Microsoft Certified: Power BI Data Analyst Associate, Tableau Desktop Specialist, HackerRank Advanced SQL, and Databricks Generative AI Fundamentals.
- 💡 Core Focus: Autonomous Agent Workflows (LangGraph, Anthropic SDK), Low-Latency Vector Databases (pgvector <120ms), Serverless Inference Optimization (ONNX Lambda <100ms), and Quantitative BI Analytics.
| Package | Description | Badges |
|---|---|---|
wranglefy |
Automated tabular data preprocessing, schema validation, and missing data auditing across CSV, JSON, Excel, and Parquet. |   |
tprdbreader |
High-performance custom byte-stream parser delivering an 80% speedup in dataset loading for research pipelines. |   |
| Venue | Paper Title | Core Contribution |
|---|---|---|
| IEEE CAI 2026 | Retrieval-Augmented AI Feedback in STEM Education | RAG evaluation pipelines, metacognitive prompting, and factual grounding |
| MDPI AI 2026 | Mapping the LLM Landscape: Cross-Family Survey | Cross-family survey across 100+ LLMs, RLHF, DPO, and benchmark evals |
| SKASE 2025 | The Behavioural Translation Style Space | Temporal cognitive dynamics & RL agent behavioral trajectory simulation |
| TT8 2026 conference | Assessing Pause Thresholds in Process Research | Parametric & non-parametric latency distribution modeling on HCI data |
| Preprint 2026 | Large Language Models: Architectures & Alignment | Attention formulations, state-space models, PEFT/LoRA, and alignment |
| arXiv 2025 | From Representation to Enactment: ABC Framework | Cognitive state transitions and mental representation modeling in NLP |
- Live Platform: multi-tenant-saas-platform-with-an-ai.onrender.com
- Engineered an adaptive LangGraph Self-RAG state graph combining hybrid retrieval (pgvector + Postgres FTS via RRF) with local DeBERTa-v3 NLI cross-encoders (~20–40ms CPU latency) for zero-cost hallucination verification and dynamic online search fallback.
- Built a production Model Context Protocol (FastMCP) server and autonomous ReAct agent over
stdio, exposing tenant-isolated vector search and Self-RAG tools directly to Claude Desktop and Cursor IDE with zero unauthorized cross-boundary leakage.
- Live Demo: deepalign-rl.streamlit.app
- Implemented DeepSeek-style Group Relative Policy Optimization (GRPO) and DPO from scratch in PyTorch, standardizing group-wise advantages to eliminate the critic model and cut peak training VRAM by ~50%.
- Engineered a zero-noise SymPy symbolic reward verifier and AST sandbox for mathematical equivalence and Chain-of-Thought format compliance, achieving 74.8% GSM8K / 49.3% MATH accuracy across automated ablation suites.
- Live Demo: flowmatch-compress.streamlit.app
- Implemented Optimal Transport Flow Matching (OT-CFM) and DiT with AdaLN-Zero from scratch in PyTorch, developing two-stage distillation (CFG + progressive trajectory consistency) to eliminate dual-pass evaluations and compress 50 ODE steps into 4 (22.05× inference speedup).
- Engineered a modular Post-Training Quantization (PTQ) suite supporting IEEE FP8 (E4M3), per-channel INT8, and grouped INT4 with activation calibration, slashing model footprint by up to 56.3% with minimal discretization drift.
- Integrated an AutoencoderKL VAE latent pipeline (
$4 \times 32 \times 32$ latent space) for high-resolution$256 \times 256$ generation, validated against photographic benchmarks and visual ODE trajectory profilers.
- Built a hybrid deep learning ensemble (Attention-LSTM + Ridge Regression) achieving 75% directional accuracy.
- Exported the PyTorch computational graph to ONNX Runtime and deployed on serverless AWS Lambda (<100ms cold-start latency).
- Live Streamlit App: olist-ecommerce-intelligence.streamlit.app
- Executed causal econometric analysis on 96,000+ customer records ($15.4M GMV) and built RFM behavioral segmentation in PostgreSQL & Streamlit.
- Live Tableau Dashboard: Public Tableau Visualization
- Engineered quantitative data models across 50,000+ records, evaluating diagnostics via AUC-ROC, KS statistic, and Gini coefficient.
- How to Stop AI Agents from Faking Task Completion.
- A Practical Guide to T-SQL
- Every Microsoft PL-300: Power BI Data Analyst Associate certification Trap in One Article, Plus the…
- Stop Paying for Tokens: Run Private LLMs on Your Laptop with Docker
- Beyond the Curve: A Holistic Guide to Activation Functions (From Sigmoid to SwiGLU)