work.html

<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Production Work | Sergey Orlov</title>
<meta name="description" content="Production case studies in credit risk ML, document intelligence, analytics engineering, and entity resolution. Engineering responsibilities and public demonstrations.">
<meta name="theme-color" content="#192a30">
<link rel="canonical" href="https://orlovtsu.github.io/work.html">
<meta property="og:type" content="website">
<meta property="og:title" content="Production Work | Sergey Orlov">
<meta property="og:description" content="Production case studies in credit risk ML, document intelligence, analytics engineering, and entity resolution. Engineering responsibilities and public demonstrations.">
<meta property="og:url" content="https://orlovtsu.github.io/work.html">
<meta property="og:image" content="https://orlovtsu.github.io/assets/portfolio/social-card.png">
<meta name="twitter:card" content="summary_large_image">
<link rel="icon" type="image/svg+xml" href="assets/portfolio/favicon.svg">
<link rel="stylesheet" href="assets/portfolio/site.css">
<script type="application/ld+json">{"@context": "https://schema.org", "@type": "Person", "name": "Sergey Orlov", "url": "https://orlovtsu.github.io/", "jobTitle": "Data Scientist", "sameAs": ["https://github.com/orlovtsu", "https://www.linkedin.com/in/orlovtsu/"], "address": {"@type": "PostalAddress", "addressLocality": "Calgary", "addressRegion": "Alberta", "addressCountry": "CA"}}</script>
</head>
<body>
<a class="skip" href="#main">Skip to content</a>
<header class="site-header">
<div class="wrap nav">
<a class="brand" href="index.html" aria-label="Sergey Orlov — home">
<span class="monogram" aria-hidden="true">SO</span> Sergey Orlov</a>
<nav class="nav-links" aria-label="Main navigation">
<a href="index.html#work">Work</a>
<a href="index.html#projects">Open source</a>
<a href="index.html#about">About</a>
<a class="nav-contact" href="mailto:orlovtsu@gmail.com">Let’s talk <span aria-hidden="true">↗</span>
</a>
</nav>
</div>
</header>
<main id="main">
<div class="wrap">
<section class="page-intro">
<a class="back" href="index.html">← Back to portfolio</a>
<div class="eyebrow">Selected production work</div>
<h1>Ownership beyond<br>the model.</h1>
<p class="lead">Selected systems I build and own: credit decisioning, document intelligence, analytics infrastructure, and entity resolution.</p>
<div class="actions">
<a class="button" href="downloads/Sergey_Orlov_Decision_Engine_Case_Study.pdf">Architecture overview <span class="small">PDF ↓</span>
</a>
<a class="button secondary" href="downloads/Sergey_Orlov_Resume.pdf">Résumé <span class="small">PDF ↓</span>
</a>
</div>
</section>
<article class="case" id="decision-engine">
<div class="case-heading"><div><div class="eyebrow">01 / Cashco Financial · 2024–present</div><h2>Credit decisioning,<br>from design to operation.</h2></div><p><strong>My role: architect and end-to-end technical owner.</strong><br>From financial-data ingestion and model development to production delivery, evaluation, and ongoing operation.</p></div>
<p class="lead">I architected and own a production decisioning system that combines financial-data ingestion, transaction and behavioral feature engineering, calibrated ML scoring, explicit policy orchestration, explainability, versioning, and production monitoring.</p>
<div class="case-columns">
<div><h3>Modeling</h3><p>Productionized routed, calibrated risk models and engineered features from financial transactions and behavior. Evaluate discrimination and probability quality, and use champion/challenger evaluation to assess changes.</p></div>
<div><h3>Decisioning</h3><p>Keep model estimation separate from policy decisions, with explicit review paths and safe fallbacks. Integrate SHAP-based explanations and reason codes with versioned decision evidence to support review and investigation.</p></div>
<div><h3>Production</h3><p>Build and operate Python/FastAPI services using Docker and Azure. CI/CD, regression validation, versioned releases, and structured logging support controlled delivery and auditable decisions.</p></div>
<div><h3>Lifecycle</h3><p>Use monitoring, shadow evaluation, and outcome feedback to guide model and policy iteration. Connect investigation of input quality and model behavior with the next round of evaluation and engineering work.</p></div>
</div>
<p class="case-note">The architecture overview explains my engineering responsibilities through a conceptual flow. The linked labs are independent synthetic demonstrations; they do not reproduce employer configurations, decision rules, or performance results.</p>
<div class="actions"><a class="text-link" href="downloads/Sergey_Orlov_Decision_Engine_Case_Study.pdf">Read the architecture overview (PDF) ↓</a><a class="text-link" href="https://github.com/orlovtsu/open-decisioning-lab">Inspect the decisioning lab ↗</a><a class="text-link" href="https://github.com/orlovtsu/default-risk-lifecycle-lab">Inspect the lifecycle lab ↗</a></div>
</article>
<article class="case" id="document-ai">
<div class="case-heading"><div><div class="eyebrow">02 / Cashco Financial · Document intelligence</div><h2>Make unstructured data<br>useful and reviewable.</h2></div><p><strong>My role: design and implementation.</strong><br>Document extraction, validation, and quality gates for financial inputs.</p></div>
<div class="case-columns"><div><h3>My contribution</h3><p>I build document-processing workflows that turn unstructured financial inputs into usable data. Extraction is paired with explicit quality gates, financial-consistency checks, and review paths for uncertain or incomplete evidence.</p></div><div><h3>Explore the general methods</h3><p>My independent public labs illustrate extraction evaluation, data-quality checks, and transaction feature engineering using synthetic inputs. They make the methods inspectable without presenting employer data or operational results.</p></div></div>
<div class="actions"><a class="text-link" href="https://github.com/orlovtsu/document-ai-reliability-lab">Inspect the synthetic reliability lab ↗</a><a class="text-link" href="https://github.com/orlovtsu/cashflow-intelligence-lab">Inspect the cash-flow prototype ↗</a></div>
</article>
<article class="case" id="data-platforms">
<div class="case-heading"><div><div class="eyebrow">03 / Cashco Financial · Data engineering</div><h2>Build the foundations<br>for useful analytics.</h2></div><p><strong>My role: platform development and operation.</strong><br>Airflow orchestration, dbt modeling, data quality, and Metabase reporting.</p></div>
<div class="case-columns"><div><h3>My contribution</h3><p>I architected an analytics platform using Airflow, dbt, and Metabase. My ownership spans ingestion, orchestration, analytical modeling, data quality, reporting, and ongoing operation.</p><p>The engineering work connects maintainable pipelines with curated data models and reporting that business teams can use and review.</p></div><div><h3>Public engineering examples</h3><p>The accompanying repository contains generic Airflow, dbt, PostgreSQL, and reverse-proxy examples. Its design notes discuss common trade-offs in ingestion, data quality, and BI engineering.</p><p>These are illustrative materials informed by professional experience. They do not document an employer’s current deployment, infrastructure capacity, performance, costs, or security posture.</p></div></div>
<div class="actions"><a class="text-link" href="https://github.com/orlovtsu/analytics-platform-portfolio">Explore the generic examples ↗</a><a class="text-link" href="https://github.com/orlovtsu/analytics-platform-portfolio/blob/main/docs/architecture.md">Read the design discussion ↗</a></div>
</article>
<article class="case" id="entity-resolution">
<div class="case-heading">
<div>
<div class="eyebrow">04 / Cybera · Applied NLP</div>
<h2>Entity resolution<br>at distributed scale.</h2>
</div>
<p>Connect unstructured records with learned representations, retrieval, and reproducible distributed execution.</p>
</div>
<div class="case-columns">
<div>
<h3>What I built</h3>
<p>Designed and deployed an entity-resolution pipeline for unstructured data using transformer embeddings, Siamese neural networks, and retrieval. GPT-assisted labeling and normalization reduced manual preparation.</p>
</div>
<div>
<h3>Execution and evaluation</h3>
<p>Scaled execution with Docker, multiprocessing, GPU acceleration, and a 24-node GCP virtual-machine cluster. The reported classification accuracy was 92% on the project evaluation.</p>
<div class="tags">
<span>BERT / LLaMA2</span>
<span>FAISS</span>
<span>TensorFlow</span>
<span>GCP</span>
</div>
</div>
</div>
<p class="case-note">Results are summarized from my professional experience. Employer source code and customer data are not published. Selected earlier collaborative geospatial work is linked from the research page.</p>
</article>
</div>
</main>
<section class="contact" aria-labelledby="contact-title">
<div class="wrap contact-grid">
<div>
<div class="eyebrow light">Start a conversation</div>
<h2 id="contact-title">Let’s connect.</h2>
<p>Based in Calgary, Canada.</p>
</div>
<div class="actions">
<a class="button" href="mailto:orlovtsu@gmail.com">Get in touch <span aria-hidden="true">↗</span>
</a>
</div>
</div>
</section>
<footer class="site-footer">
<div class="wrap footer-row">
<span>© 2026 Sergey Orlov · Calgary, Canada</span>
<div class="footer-links">
<a href="https://www.linkedin.com/in/orlovtsu/">LinkedIn</a>
<a href="https://github.com/orlovtsu">GitHub</a>
<a href="downloads/Sergey_Orlov_Resume.pdf">Résumé (PDF)</a>
<a href="archive.html">Earlier research</a>
</div>
</div>
</footer>
</body>
</html>