<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Gülşah Bayhan Ara (M.Sc.) on Medium]]></title>
        <description><![CDATA[Stories by Gülşah Bayhan Ara (M.Sc.) on Medium]]></description>
        <link>https://medium.com/@gulsahbayhan?source=rss-62bba06b0d2a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*obTX_tlFlSBfEcUpXfUk0g.jpeg</url>
            <title>Stories by Gülşah Bayhan Ara (M.Sc.) on Medium</title>
            <link>https://medium.com/@gulsahbayhan?source=rss-62bba06b0d2a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 22:27:36 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@gulsahbayhan/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Agile Meets AI: Why Sprint Metrics Break with Non- Deterministic Models]]></title>
            <link>https://medium.com/@gulsahbayhan/agile-meets-ai-why-sprint-metrics-break-with-non-deterministic-models-1af45f7ccc71?source=rss-62bba06b0d2a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1af45f7ccc71</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[scrum]]></category>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Gülşah Bayhan Ara (M.Sc.)]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 23:04:50 GMT</pubDate>
            <atom:updated>2026-11-04T23:04:50.028Z</atom:updated>
            <content:encoded><![CDATA[<p>By Gülşah Bayhan Ara (M.Sc.)</p><p><em>Senior Technical Product Manager</em></p><h4><strong>The Deterministic Fallacy of Standard Agile</strong></h4><p>For over two decades, Agile frameworks — and Scrum in particular — have been built on a core, unspoken assumption: <strong>software development is largely deterministic.</strong></p><p>In standard SaaS and application development, if an engineer writes a function with defined inputs, mock data, and unit tests, the output is binary. A ticket moves from “In Progress” to “Done.” We estimate user stories using Fibonacci points, calculate velocity across two-week cadences, and draw linear burndown charts that satisfy leadership dashboards.</p><p>When engineering teams transition into building products powered by Machine Learning (ML), Deep Learning, and Large Language Models (LLMs), this deterministic foundation cracks.</p><p>Forcing AI and data science initiatives into traditional two-week Scrum iterations frequently produces synthetic velocity, developer burnout, and frustrated product leadership. The problem isn&#39;t the team’s capability, and it doesn&#39;t mean Agile is dead. The breakdown happens because standard sprint metrics were never designed for <strong>non-deterministic systems</strong>.</p><h4>Why the Mechanics Collide: Traditional Software vs. Applied AI</h4><ul><li><strong>Output Behavior</strong></li><li><em>Traditional Software:</em> Deterministic ($f(x) \rightarrow y$ reliably).</li><li><em>Applied AI &amp; Non-Deterministic Systems:</em> Probabilistic (confidence scores, hallucinations, model drift).</li><li><strong>Effort vs. Outcome</strong></li><li><em>Traditional Software:</em> Linear (more engineering hours $\approx$ feature complete).</li><li><em>Applied AI &amp; Non-Deterministic Systems:</em> Non-linear (days of hyperparameter tuning or dataset pruning may yield 0% metric gain).</li><li><strong>Task Predictability</strong></li><li><em>Traditional Software:</em> High (APIs, database schemas, UI, and CRUD logic are well-understood).</li><li><em>Applied AI &amp; Non-Deterministic Systems:</em> Low (hypothesis-driven exploration, data distribution shifts, edge-case surprises).</li><li><strong>Definition of Done (DoD)</strong></li><li><em>Traditional Software:</em> Code compiles, unit/integration tests pass, feature deploys.</li><li><em>Applied AI &amp; Non-Deterministic Systems:</em> Model achieves target evaluation benchmarks (e.g., F1 $\ge$ 0.88) without performance regression.</li><li><strong>Failure Mode</strong></li><li><em>Traditional Software:</em> Explicit crashes, syntax errors, stack traces, compiler errors.</li><li><em>Applied AI &amp; Non-Deterministic Systems:</em> Silent failures, semantic data leakage, latent degradation under distribution changes.</li></ul><h4>The Three Points of Failure in AI Sprints</h4><h4>1. The Story Point Paradox (Research Is Not CRUD)</h4><p>In standard Scrum, a 5-point story represents a known level of complexity. In applied AI, a task like <em>“Improve NER precision on domain-specific contract data”</em> might look like an 8-point ticket.</p><p>An engineer might spend three days cleaning data, retraining, and evaluating, only to discover severe label noise in the training set or that the model overfits on edge cases. At the end of the sprint, the accuracy metric has not improved — or has degraded.</p><p>Did the team deliver zero points? According to standard Jira burnups: yes. In reality, the team successfully eliminated a failed hypothesis, which is vital technical progress. Treating exploration as incomplete delivery penalizes scientific inquiry and encourages engineers to pick “safe,” shallow tasks rather than solving core architectural bottlenecks.</p><h4>2. The Elusive “Definition of Done”</h4><p>In classical software, a feature is done when acceptance criteria are satisfied and regression suites pass.</p><p>With non-deterministic models, “Done” is a moving target:</p><ul><li>A prompt pipeline might answer 95% of standard queries accurately but produce unacceptable hallucinations on 5% of tail queries.</li><li>Fine-tuning an open-source LLM for latency might degrade reasoning capability on structured JSON extraction.</li></ul><p>If your DoD requires absolute predictability, an AI story can stay trapped in the sprint backlog indefinitely.</p><h4>3. Continuous Integration vs. Continuous Retraining</h4><p>Traditional CI/CD pipelines validate deterministic syntax and integration contracts. In AI-native products, pipelines must account for data validation, model drift, inference latency under variable token lengths, and stochastic outputs. Agile cadences rarely account for the infrastructure dependencies inherent to MLOps.</p><h3>The Solution: A Dual-Track Delivery Framework for AI Products</h3><p>To manage non-deterministic AI initiatives without sacrificing accountability, technical leadership must decouple <strong>Hypothesis Discovery</strong> from <strong>Software Delivery</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XVWePz0_Uh121lts-ZA5OQ.png" /></figure><h4>1. Replace Story Points with Timeboxed Research Spikes</h4><p>Do not ask data scientists or AI engineers to point experimental tasks with Fibonacci numbers. Instead, frame model experimentation as <strong>timeboxed research spikes</strong>:</p><ul><li>Give a dedicated allocation (e.g., 3 days or 1 sprint cycle).</li><li>Define upfront the go/no-go evaluation threshold: <em>“Explore fine-tuning Llama-3–8B on customer support logs. The goal is achieving an F1 score $\ge 0.85$ on the benchmark eval set. If not viable within 4 days, document findings and pivot to RAG optimization.”</em></li><li>Success is defined by the <strong>clarity of the finding</strong>, not whether code merged into production.</li></ul><h4>2. Statistical Definitions of Done (Eval-Driven Development)</h4><p>Replace qualitative acceptance criteria with rigorous evaluation benchmarks (Evals):</p><ul><li>Establish fixed golden evaluation datasets before writing prompts or training models.</li><li>Define DoD statistically: <em>“Model passes when semantic drift remains under 3% and latency stays $&lt; 450\text{ms}$ at p95 across the test suite.”</em></li></ul><h4>3. Treat Prompts and Models as Dependencies, Not Features</h4><p>The user-facing product, API wrappers, caching layers, safety guardrails, and telemetry should run on predictable, iterative Agile sprints. The underlying stochastic model should be treated as a versioned dependency. When Track 1 validates a new model checkpoint, Track 2 integrates and deploys it through standard CI/CD pipelines.</p><h3>The Role of the AI Product Leader</h3><p>Navigating non-deterministic products requires a shift in leadership mindset. Managing an AI portfolio is not about enforcing Jira orthodoxy; it is about building an environment where experimentation is structured, technical debt is actively managed, and business stakeholders understand probabilistic risk.</p><p>Agile is not obsolete in the era of artificial intelligence — but dogmatic Scrum is. By adapting our metrics to respect the non-deterministic nature of machine intelligence, we bridge the gap between scientific innovation and reliable product delivery.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1af45f7ccc71" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>