Maxim Kleyer — GenAI / AI Engineer

I build AI systems — and the harnesses that prove they work.

Generative AI engineer (CS + Math, Macaulay Honors) with professional RAG experience and a habit of validating everything: my college football projection system is backtested over 7,401 games with calibrated win probabilities — landing within a point of Vegas closing lines.

College Football Projection System

Weekly game projections for all FBS college football: set a week, run one command, get every game with projected scores, spread, total, win probability, and comparison against market lines. Python · LightGBM · pandas · Parquet · 17-endpoint REST ingestion · automated test suite.

7,401FBS games, walk-forward validated (2017–2025)

75.5%straight-up accuracy — Vegas closing lines: 76.3%

16.37margin RMSE — vs 23.06 always-pick-home, 18.75 SP+ differential

0.162Brier score, isotonic-calibrated, reported out-of-fold

Built like software, judged like a hypothesis

A ridge-regression rating layer anchored on preseason priors, two LightGBM models over ~28 features, a 10,000-run Monte Carlo season simulator with empirically measured correlated team-strength uncertainty, and an experiment lab where a candidate idea must beat the baseline on both fixed walk-forward folds to be adopted. A weekly loop projects the slate, grades the results, and measures drift against a matched baseline — drift is a number rather than a feeling.

What the model can’t do, measured

Across seven spread buckets and seven total buckets, no betting edge is statistically significant. Hit rates run 48.6%–55.5% against the 52.4% break-even, non-monotonically — exactly what chance produces across seven looks. The system says so on its own report card.

Building the significance test that kills your own headline is the point: a model you can’t trust to admit failure can’t be trusted to claim success.

Read the full case study → GitHub repo: being published — github.com/MaximKleyer

Projects & professional GenAI

NFL Game Projection Model

Python · pandas · NumPy

Bottom-up NFL projections built from snap-share rosters with per-team calibration, plus a walk-forward evaluation harness across 816 games (2023–2025): 63.8% straight-up accuracy, 10.40 margin MAE, with error analysis used to calibrate systematic bias to approximately zero. The methodological predecessor to the CFB system — where the evaluation-first habit started.

Code →

RGB Heart Rate Monitor

Python · FFT · Signal processing — Undergraduate thesis

Measures heart rate from live camera input alone: isolates the heartbeat signal from red, green, and blue channel variation using Fast Fourier Transform and noise-reduction filtering. Carried solo from problem definition through working system, written thesis, and defense at Macaulay Honors College.

Live demo → Code →

RAG Pipeline — SpruceTech

Python · LlamaIndex — AI development internship

Built a retrieval-augmented generation pipeline that grounded LLM output in retrieved source material to score free-text student answers consistently — prompt iteration and retrieval grounding turning unreliable generative output into repeatable, reviewable results behind a user-facing UI.

Employer work — no public repo

Pocket Summoner

React/Vite SPA live on GitHub Pages: combat and quest engine over a 191-entity database, with Python scripts generating and validating the dataset.

Play →

Valorant GM Simulator

React/Vite simulation: season/offseason engine, player development, age-based retirement, and a multi-phase competitive calendar, iterated across releases.

Play →

Both built with AI-assisted development used with discipline: tested, reviewed, shipped.

Where the delivery muscle comes from

A year of professional DevOps/QA engineering means models get treated like software: tested, deployed, monitored.

Mar 2025 — Present

Technical Support Consultant

C&P Management · New York, NY

Diagnoses and resolves application and network failures across 1,000+ FDNY workstations, analyzing logs and live system state in real time
Triages a continuous ServiceNow queue by business impact with systematic root-cause analysis
Primary stakeholder contact during incidents — translating technical root causes into clear updates for non-technical audiences

Feb 2024 — Jan 2025

DevOps Engineer & QA Automation

NuArch · New York, NY

Designed and maintained Jenkins/Git CI/CD pipelines automating build, test, and deployment across dev, QA, test, and UAT environments
Built automated test suites (Selenium, Cucumber) gating releases on unit, integration, and regression checks; wrote Cucumber (Gherkin) scenarios translating business requirements into testable acceptance criteria
Python and Bash automation eliminating repetitive manual work; documented runbooks

Jun 2023 — Aug 2023

AI Development Intern

SpruceTech · Clifton, NJ

Built the Python + LlamaIndex retrieval-augmented generation pipeline described above — retrieval grounding and prompt iteration making generative scoring consistent, repeatable, and reviewable

Tools that get used, not listed

GenAI / LLM

RAGLlamaIndexRetrieval grounding Prompt engineeringStructured outputsNLP AI-assisted development (Claude Code)

ML

PythonLightGBMscikit-learn PyTorchRidge regressionIsotonic calibration Monte Carlo simulationWalk-forward validation Drift measurementModel evaluation pandasNumPy

Delivery

JenkinsGitHub ActionsGit CI/CD designSeleniumCucumber Automated quality gatesMonitoring & log analysis

Data / Engineering

REST API ingestionParquet caching Data validation & reconciliationSQL JSONLinux CLI

Cloud / Web

AWSMicrosoft Azure ReactVite

Languages

PythonSQLJavaScript JavaBashPowerShell

Foundation

B.A., Computer Science & Mathematics (dual major)

Macaulay Honors College at CUNY · New York, NY · May 2024 · GPA 3.7

Certifications

Generative AI Fundamentals — Databricks Academy (2025)
Applied Generative AI Specialization — Simplilearn / Microsoft Azure AI (in progress)