Brian Demajh

About

I'm an R&D-focused AI/ML engineer building innovative, scalable solutions for high-stakes environments. My work spans government and defense applications, early-stage startups, and Fortune-100 enterprises.

My PhD research at UC Santa Barbara focused on deep reinforcement learning for closed-loop control of neural systems — building an end-to-end RL stack that ingests synchronized point-clouds of neural signals (microarray, EEG, fMRI) and learns guidance policies for both model-based and model-free control of neural subsystems.

Research Expertise

Deep RL (policy gradients, actor-critic, curriculum learning), multi-sensor fusion (lidar-vision-IMU), trajectory optimization, Bayesian filtering, computer vision, synthetic data generation, and numerical optimal control.

Technical Skills

Python Go PyTorch TensorFlow CUDA OpenCV ROS MuJoCo AWS GCP Docker Kubernetes Terraform Snowflake Spark Faiss Neo4j

Experience

Automated GTM-Planning Platform

Conceived and shipped a platform aggregating multi-modal data (CRM, intent, OSINT) with real-time account-based playbooks. Built entity-resolution and graph-linking models (PyTorch + Faiss) enabling <500ms semantic search via a proprietary company embedding model.
Vision-Based Process Mining

Deployed a CV pipeline tracking operator-machine interactions across four factories. Reduced safety incidents by 43% and saved $2.1M/yr. Process models guided a $100M reshoring effort to build two new US factories. IP became the basis of a $6M seed-funded company.
Identity Verification & Fraud Detection

Built Mixture-of-Experts CV models outperforming AWS Rekognition on face-match tasks. Developed a synthetic fake-ID generator for face-swap detection training. Models check >50K jobs daily for fraud in production.
Real-Time Churn Prediction

Unified 7 SaaS data streams into a Snowflake lake-house with XGBoost + SHAP explainability. System serves 150+ mid-market customers in production.
Compliance QA Chatbot

Built a transformer-based QA model fine-tuned on 100K+ security questionnaire pairs, producing instant risk-assessment narratives. Product raised $4M Seed.

Testimonials

Brian Demajh is an exceptional data and applied ML scientist and leader. He consistently converts complex projects and technical challenges/tasks into manageable action plans. His recommendations are concise, data-driven, thorough and immediately actionable. Brian integrates quickly and well with existing teams, elevating their performance and knowledge. Our data science team's capability expanded substantially the day after he joined us. He is methodical, approaches problems rigorously, and often gets ahead of schedule. Working with Brian is straightforward—no surprises, no spin, pure directness. He earns trust quickly, then reinforces it with concrete results. His calm demeanor steadies in high-pressure situations, his presence, determination, resilience and depth of expertise significantly lowers project risk.

— Neal Oman, CTO Vouched.id

Brian Demajh stands out in a field crowded with pretenders—he actually delivers working solutions that create real business value. He navigates ambiguous problems with confidence, doesn't get paralyzed by incomplete data, and consistently drives projects to completion without needing constant direction. His combination of rigorous credentials and hands-on problem-solving ability makes him invaluable to any team tackling complex data challenges.

Read full version

Brian Demajh stands out in a field crowded with theoretical knowledge but lacking practical execution. What impressed me most is his ability to navigate ambiguous problems with confidence. We challenged him with a particularly complex NLP issue and he was very successful in developing a solution. Brian doesn't get paralyzed by incomplete data or shifting requirements, but instead develops robust approaches that account for uncertainty. He consistently identifies opportunities and drives projects forward without needing constant direction, yet never leaves work half-finished. While many data scientists can discuss algorithms eloquently, Brian actually delivers working solutions that create real business value. His combination of rigorous academic training and hands-on problem-solving ability makes him invaluable to any team tackling complex data challenges.

— Justin Beals, CEO StrikeGraph

I've worked with him in many different contexts and can recommend him without reservation. He's capable of huge lifts while also mentoring your staff engineers to accomplish more themselves.

— Jay Bartot, CTO AirSignal, Former CTO Madrona Labs

Education

Ph.D., Computer Science

UC Santa Barbara · 2015 – 2019

Dissertation: Deep Reinforcement Learning for Control of Neural Systems. Created an end-to-end RL stack that ingests synchronized point-clouds of neural signals (microarray, EEG, fMRI) and learns closed-loop guidance policies for model-based and model-free control of neural subsystems. I was the first person to show that neural systems (from cellular to cortical scales) could be controlled with deep reinforcement learning.

M.S., Computer Science

University of Chicago · 2013 – 2014

Focus: Machine Learning.

B.S., Neuroscience & Mathematics

Johns Hopkins University · 2006 – 2009

Writing

Blog

Long-form articles on AI/ML engineering, data science value generation, visual RAG, reinforcement learning, and building production systems.

Read the blog →

Dossier

90+ concise Q&A pages covering AI/ML research papers, techniques, companies, and tooling — curated daily.

Browse the dossier →

Open Source

Prestige

A deduplication-aware value store built on RocksDB. Supports exact (SHA-256) and semantic (neural embedding) dedup, TTL expiration, LRU eviction, transactions, a REST API, and Python/C++ interfaces.

View on GitHub →

Emacs AI Extensions

AI-powered coding extensions for Emacs. Includes an inline completion engine and a RAG-based QA system that indexes your codebase into a local Chroma vector store for context-aware answers via OpenAI.