Sitelet https://darthmanwe.github.io/

Senior Applied ML | GenAI | MLOps

Building production AI systems that deliver measurable outcomes.

I help teams ship reliable AI products faster by combining graph-aware LLM workflows, document AI, robust ML engineering, and cloud-native MLOps discipline. My focus is turning messy inputs into observable systems with measured delivery impact.

GitHub Profile LinkedIn Extended CV Email
Kutlu Mizrak profile photograph

Impact Highlights

99.9%

Production uptime with 100k+ daily API requests.

230%

Pipeline acceleration through workflow hardening.

35%

F1 score increase in entity extraction systems.

98.7%

OCR extraction accuracy across 5M+ lines.

Featured Projects

Selected builds with measurable behavior, explicit trade-offs, and reproducible engineering artifacts.

Graph Memory for LLM Storytelling

Lorekeeper

Read-write Neo4j memory loop for long-form story generation. The system extracts entities after each segment, retrieves graph facts before the next one, and uses pre-generation Cypher guardrails to cut contradiction scores by 61% against a rolling-context baseline.

Neo4j LangGraph ChromaDB FastAPI OpenTelemetry 137 tests
View Repository

Medical GraphRAG Benchmark

Medical GraphRAG System

Neo4j + ChromaDB medical-document RAG comparison with provenance-linked citations and five composable Cypher retrieval patterns. Graph retrieval reaches 0.96 faithfulness, 0.78 context precision after re-ranking, and roughly 3.5s generation latency.

Neo4j Cypher FastAPI Claude ChromaDB 29 tests
View Repository

Fault-Tolerant Distributed RL

Distributed RL Training Platform

Ray + PyTorch PPO system built for heterogeneous workers, checkpoint recovery, stale-rollout rejection, and worker churn. It ships with 42 passing tests, zero lint/type errors, and validated CartPole, scheduling, and churn training runs.

Ray PyTorch PPO OpenTelemetry Prometheus Grafana
View Repository

Agentic Document Automation

PDF_to_Presentation / pdfdeck

LangGraph pipeline that turns medical-textbook PDF excerpts into grounded slide decks with clean figure extraction, fidelity-checked bullets, optional Turkish translation, and a packaged Windows app. It reduces a 3,214-fragment extraction failure to 8 clean figures with 0 fragments.

LangGraph Claude Vision PyMuPDF python-pptx PyInstaller 91 tests
View Repository

Geospatial VLM Review Funnel

SatChangeGate

Cost-controlled satellite change-detection triage for Sentinel-2 imagery. A deterministic quality and classical gate filters 64.6% of OSCD tiles before VLM review, then gated verification reaches 0.971 sampled precision on the held-out split.

Sentinel-2 OSCD Claude VLM Pydantic Docker 112 tests
View Repository

Cross-League Basketball Translation

HoopsLab

Inference system for estimating how basketball production translates across the EuroLeague, NBA, and G League. Built on 414 real transfers and 22,297 player-seasons, it publishes 80% intervals, selection limits, and misses up front: usage-rate MAE reaches 0.0332 and beats the league-mean baseline by 22.4%, while true shooting loses and is reported that way.

Python Polars statsmodels Cloudflare Workers D1 374 tests
View Repository

NBA Lineup Fit Forecaster

LineupIQ

Live Cloudflare Worker for scoring any five NBA players, including groups that have never shared the floor. It reconstructs 698,314 shots into lineup context, serves a closed-form shot-selection model that matches the Python fit to 1e-9, and uses a 422 INSUFFICIENT_SUPPORT contract instead of inventing digits for the 99% of lineups without enough possessions.

Python TypeScript Cloudflare Workers Hono Next.js 341 tests
View Repository

Career Highlights

Lore Machine

Senior AI/ML Engineer; Director of AI/ML Operations

Remote, CA | Jun 2023-Mar 2026

At Lore Machine, I owned the LLM automation backbone from architecture through production operations, turning unstructured story inputs into validated entities, retrieval-ready knowledge, and product workflows that could be audited and reused. The work blended agentic orchestration, GraphRAG, safety controls, and production reliability, sustaining 99.9% uptime across services handling 100,000+ daily API requests while accelerating story-visualization workflows by 230% and improving entity extraction F1 by 35%.

Empower AI (formerly NCI Inc.)

ML Engineer I-II / Team Lead / Project Consultant

Remote, Reston, VA | Sep 2020-May 2023

At Empower AI, formerly NCI Inc., I led applied ML delivery across OCR/NLP, document AI, Transformer deployment, extraction QA, and client-facing validation workflows for intelligence use cases. I helped move noisy documents from raw extraction into repeatable analyst-ready pipelines, raising extraction accuracy to 98.7%, cutting processing time by 60%, reaching 99.8% CI/CD reliability, and reducing manual review effort by 85%.

Mercyhurst University CIRAT

Data Science Consultant / Assistant Lab Director / Senior Data Science Team Lead

On-site, Erie, PA | May 2021-May 2023

At Mercyhurst CIRAT, I combined data science consulting, lab leadership, intelligence analysis, and technical mentoring while leading 25 students across 8 research and applied analytics projects. I built structure around project intake, OSINT automation, scraper reliability, and analyst handoffs, reducing kickoff time by 40%, raising lab utilization by 150%, surfacing 200+ indicators weekly, and cutting analysis time by 50%.

Core Stack

LLM and Agentic Systems

LangGraph, structured outputs, tool use, workflow orchestration, grounded generation, and validation loops.

Graph and Retrieval

Neo4j 5.x, Cypher, GraphRAG, vector retrieval, ChromaDB, provenance chains, and semantic re-ranking.

ML and Evaluation

PyTorch, Transformers, OCR/NLP extraction, RL training systems, RAG eval, and benchmark harness design.

Backend, MLOps, and Cloud

Python, FastAPI, Streamlit, Docker, Kubernetes, OpenTelemetry, Prometheus, AWS, Azure, GCP, and CI/CD.

Certifications

AWS Certified Machine Learning - Specialty Google Cloud Machine Learning Engineer Azure AI Engineer Associate