About AI, ML, LLMs and more by Yvette Tsai. https://yvettetsai22.github.io/blog/ Thu, 10 Sep 2026 13:46:05 +0000 Thu, 10 Sep 2026 13:46:05 +0000 Jekyll v3.10.0 <p>The most dangerous bug is the one that doesn’t crash.</p> <p>Today I shipped a form-panel editing feature for a forecasting chatbot. Users can now edit time slots directly from the UI without typing natural language commands — the feature routes all edits through a new <code class="language-plaintext highlighter-rouge">/update-slots</code> command that reuses the same validation guards as the chat path.</p> <p>What shipped: <code class="language-plaintext highlighter-rouge">run_slot_update()</code> extracted as a shared function, <code class="language-plaintext highlighter-rouge">UpdateSlotsCommand</code> with 12/12 tests green, and a WebSocket handler refactored from 218 down to 132 lines with <code class="language-plaintext highlighter-rouge">_handle_form_edit</code> deleted entirely.</p> <p>At every step, bugs that didn’t fail... Thu, 25 Jun 2026 08:00:00 +0000 https://yvettetsai22.github.io/blog/everything-returned-true-debugging-silent-failures/ https://yvettetsai22.github.io/blog/everything-returned-true-debugging-silent-failures/ python debugging ai tutorial code <h2 id="bug-1-applyreplace-always-returned-true">Bug 1: <code class="language-plaintext highlighter-rouge">_apply_replace</code> always returned <code class="language-plaintext highlighter-rouge">True</code></h2> <p><code class="language-plaintext highlighter-rouge">_merge_slot_entries</code> called <code class="language-plaintext highlighter-rouge">_apply_replace</code> to detect whether a slot value had changed. The function compared nothing. It returned <code class="language-plaintext highlighter-rouge">True</code> unconditionally.</p> <p>This meant the “no changes detected” early-return in <code class="language-plaintext highlighter-rouge">run_slot_update</code> never fired — any update, even one with identical values, passed through as a real change. Tests that expected the guard to catch no-ops were passing for the wrong reason.</p> <div class="language-python highlighter-rouge"><div class="highlight"><pre class="highlight"><code><span class="c1"># Before: unconditional return </span><span class="k">def</span> <span class="nf">_apply_replace</span><span class="p">(</span><span class="n">slot</span><span class="p">,</span> <span class="n">new_value</span><span class="p">):</span> <span class="n">slot</span><span class="p">.</span><span class="n">value</span>... Thu, 25 Jun 2026 08:00:00 +0000 https://yvettetsai22.github.io/blog/everything-returned-true-debugging-silent-failures/2/ https://yvettetsai22.github.io/blog/everything-returned-true-debugging-silent-failures/2/ python debugging ai tutorial code <p>The most dangerous bug is the one that doesn’t crash.</p> <p>Today I shipped a form-panel editing feature for a forecasting chatbot. Users can now edit time slots directly from the UI without typing natural language commands — the feature routes all edits through a new <code class="language-plaintext highlighter-rouge">/update-slots</code> command that reuses the same validation guards as the chat path.</p> <p>What shipped: <code class="language-plaintext highlighter-rouge">run_slot_update()</code> extracted as a shared function, <code class="language-plaintext highlighter-rouge">UpdateSlotsCommand</code> with 12/12 tests green, and a WebSocket handler refactored from 218 down to 132 lines with <code class="language-plaintext highlighter-rouge">_handle_form_edit</code> deleted entirely.</p> <p>At every step, bugs that didn’t fail... Thu, 25 Jun 2026 08:00:00 +0000 https://yvettetsai22.github.io/blog/everything-returned-true-debugging-silent-failures/view-all/ https://yvettetsai22.github.io/blog/everything-returned-true-debugging-silent-failures/view-all/ python debugging ai tutorial code <p>Large Language Models (LLMs) are rapidly transforming how we interact with technology. But with great power comes great… complexity. Understanding what’s happening inside these intricate models is crucial for performance optimization, debugging, and ensuring responsible AI practices. That’s where observability comes in.</p> <p>While automated observability tools offer a quick start, sometimes you need to get your hands dirty. You need granular control over what data you collect and how you analyze it. That’s where manual instrumentation shines. In this post, we’ll explore how to leverage manual instrumentation with Google’s ADK (Application Development Kit) to gain deeper insights into your LLM... Sun, 22 Feb 2026 16:40:15 +0000 https://yvettetsai22.github.io/blog/level-up-your-llm-observability-a-hands-on-guide-to-manual-instrumentation-with-google-adk/ https://yvettetsai22.github.io/blog/level-up-your-llm-observability-a-hands-on-guide-to-manual-instrumentation-with-google-adk/ blog <p>Bing Custom Search API lets you build a focused search engine that returns results only from domains you trust. Instead of the entire web, you define a curated set of sites. This sharpens relevance, improves compliance, and makes answers faster and cleaner for your users. It’s ideal for internal knowledge portals, customer-facing help centers, and vertical search experiences where control matters more than breadth.</p> <p>Compared to general Bing Web Search, you get domain scoping, ranking controls, pinned results, and filters you can tune without code. You’ll create two things:</p> <ol> <li>a <strong>Custom Bing Search instance</strong>,<br /> where you configure domains... Wed, 30 Aug 2023 12:05:21 +0000 https://yvettetsai22.github.io/blog/bing-custom-search-api/ https://yvettetsai22.github.io/blog/bing-custom-search-api/ blog blog