<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0">
  <channel>
    <title>Tathastha Labs Insights</title>
    <link>https://tathastha.com/insights</link>
    <description>Practical healthcare AI, with our verdict and the evidence behind it.</description>
    <item>
      <title>NER: general vs fine-tuned, and when to switch</title>
      <link>https://tathastha.com/insights/fine-tuning/ner-when-a-general-model-is-enough-and-when-to-fine-tune</link>
      <guid>https://tathastha.com/insights/fine-tuning/ner-when-a-general-model-is-enough-and-when-to-fine-tune</guid>
      <pubDate>Mon, 16 Nov 2026 00:00:00 GMT</pubDate>
      <description>Named entity recognition explained with a worked clinical example: what general models miss, when to fine-tune, and why finding an entity is only half the job.</description>
    </item>
    <item>
      <title>RAG reranking: when a second pass pays off</title>
      <link>https://tathastha.com/insights/rag/rag-reranking-when-a-second-pass-pays-off</link>
      <guid>https://tathastha.com/insights/rag/rag-reranking-when-a-second-pass-pays-off</guid>
      <pubDate>Mon, 16 Nov 2026 00:00:00 GMT</pubDate>
      <description>What a cross-encoder reranker adds to retrieval-augmented generation, how our in-browser demo works, and why you should measure before keeping it.</description>
    </item>
    <item>
      <title>Small LLMs on-device: what a 0.5B model can do</title>
      <link>https://tathastha.com/insights/llm/small-llms-on-device-what-a-0-5b-model-can-and-cant-do</link>
      <guid>https://tathastha.com/insights/llm/small-llms-on-device-what-a-0-5b-model-can-and-cant-do</guid>
      <pubDate>Mon, 16 Nov 2026 00:00:00 GMT</pubDate>
      <description>What we learned running a 0.5B language model in the browser for note drafting, receipts and RAG answers: where it works, where it doesn't, and why.</description>
    </item>
    <item>
      <title>Small fine-tuned models vs per-chart LLM calls</title>
      <link>https://tathastha.com/insights/fine-tuning/small-models-vs-per-chart-llm-calls</link>
      <guid>https://tathastha.com/insights/fine-tuning/small-models-vs-per-chart-llm-calls</guid>
      <pubDate>Mon, 16 Nov 2026 00:00:00 GMT</pubDate>
      <description>We replaced a cloud LLM pipeline for HCC coding with small fine-tuned models: about 20 minutes to about 1 minute per chart, $0.31 to $0, and better recall.</description>
    </item>
  </channel>
</rss>
