<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Marcos Alvarez</title>
    <description>The latest articles on DEV Community by Marcos Alvarez (@marcos_alvarez_251b1536b2).</description>
    <link>https://dev.to/marcos_alvarez_251b1536b2</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4209610%2F40d58f1c-1079-4e0a-b632-d969409ab01d.webp</url>
      <title>DEV Community: Marcos Alvarez</title>
      <link>https://dev.to/marcos_alvarez_251b1536b2</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/marcos_alvarez_251b1536b2"/>
    <language>en</language>
    <item>
      <title>What if silence could save your harvest? (I am building BeeBuzz, a bee classifier)</title>
      <dc:creator>Marcos Alvarez</dc:creator>
      <pubDate>Mon, 16 Nov 2026 18:34:26 +0000</pubDate>
      <link>https://dev.to/marcos_alvarez_251b1536b2/what-if-silence-could-save-your-harvest-i-am-building-beebuzz-a-bee-classifier-49nn</link>
      <guid>https://dev.to/marcos_alvarez_251b1536b2/what-if-silence-could-save-your-harvest-i-am-building-beebuzz-a-bee-classifier-49nn</guid>
      <description>&lt;p&gt;I'm building BeeBuzz for ML Zoomcamp 2026 and it's still rough. The idea: every bee species has a characteristic buzz. If a species goes quiet in an area, you could notice weeks before the flowers stay empty — and still have time to move hives and change the plan.&lt;/p&gt;

&lt;p&gt;Right now it takes short lab buzzes and gives back species and confidence in seconds. I'm working with 200 Mendeley lab recordings in total (5 species), holding out 40 to check while I keep iterating: 80% accuracy on the holdout, with the weakest species (Centris cinza, F1 0.57) clearly needing more data.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4rw600v2wlsakkqz3ssm.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F4rw600v2wlsakkqz3ssm.png" alt="BeeBuzz approach" width="800" height="533"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Two things I want to be upfront about: confidence scores run overconfident (they're vote shares, not calibrated probabilities), and the buzz-type classifier scores 100% only because the filename prefix predicts the label — a recording-setup artifact, not biology. Lab recordings, no field noise yet. Field validation is next.&lt;/p&gt;

&lt;p&gt;You can try the loop live: upload a wav, see waveform, frequency and prediction right away.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpnrptaiuvzevrg9u10f0.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpnrptaiuvzevrg9u10f0.png" alt="Live audio result" width="800" height="415"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Repo (work in progress): &lt;a href="https://github.com/marcos-Nieves-24/beebuzz" rel="noopener noreferrer"&gt;https://github.com/marcos-Nieves-24/beebuzz&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Live demo: &lt;a href="https://beebuzz-csqzda6ff5wtdlwicwuzcj.streamlit.app/" rel="noopener noreferrer"&gt;https://beebuzz-csqzda6ff5wtdlwicwuzcj.streamlit.app/&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Thanks to Alexey Grigorev and the DataTalks.Club team for the structure and community. Learning in public — feedback welcome.&lt;/p&gt;

</description>
      <category>machinelearning</category>
      <category>python</category>
      <category>datascience</category>
      <category>ai</category>
    </item>
    <item>
      <title>Deploying a Frozen ML Model with FastAPI and Docker</title>
      <dc:creator>Marcos Alvarez</dc:creator>
      <pubDate>Mon, 26 Oct 2026 22:46:31 +0000</pubDate>
      <link>https://dev.to/marcos_alvarez_251b1536b2/deploying-a-frozen-ml-model-with-fastapi-and-docker-81k</link>
      <guid>https://dev.to/marcos_alvarez_251b1536b2/deploying-a-frozen-ml-model-with-fastapi-and-docker-81k</guid>
      <description>&lt;p&gt;Module 5 of ML Zoomcamp hands you a frozen model and asks you to serve it. Sounds trivial until your probabilities are slightly off and nothing errors out.&lt;/p&gt;

&lt;p&gt;Here is the setup that worked for me, with the actual code.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Verify the artifact first
Before touching anything, check you have the right model:
sha256sum pipeline.bin&lt;/li&gt;
&lt;/ol&gt;

&lt;h1&gt;
  
  
  expected: 1646bbdcd38d4f044da6b630c5b332c93a314245a8c21929011c42de51f629f1
&lt;/h1&gt;

&lt;p&gt;I learned this the hard way. Debugging numbers on the wrong artifact wastes an afternoon.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;&lt;p&gt;Pin the environment&lt;br&gt;
uv sync --locked&lt;br&gt;
That --locked flag is the whole story behind my probability mismatch. Without it, uv resolved a different sklearn build and everything shifted slightly. The reference env pins scikit-learn==1.7.2. Don't regenerate the lockfile for the graded run.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;Load the model the same way the API does&lt;/p&gt;&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;import pickle&lt;br&gt;
from model import normalize_record&lt;/p&gt;

&lt;p&gt;with open("pipeline.bin", "rb") as f:&lt;br&gt;
    pipeline = pickle.load(f)  # DictVectorizer + LogisticRegression&lt;/p&gt;

&lt;p&gt;lead = {&lt;br&gt;
    "lead_source": "paid_ads",&lt;br&gt;
    "industry": "technology",&lt;br&gt;
    "employment_status": "employed",&lt;br&gt;
    "location": "north_america",&lt;br&gt;
    "number_of_courses_viewed": 2,&lt;br&gt;
    "annual_income": 79276.0,&lt;br&gt;
    "interaction_count": 4,&lt;br&gt;
    "lead_score": 0.41,&lt;br&gt;
}&lt;br&gt;
record = normalize_record(lead)&lt;br&gt;
prob = float(pipeline.predict_proba([record])[0, 1])&lt;br&gt;
print(round(prob, 3))  # 0.533&lt;br&gt;
The key line is normalize_record. It applies the training-set medians and "NA" for missing categoricals. Hand-rolling imputation gives you a different vector and a different answer.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Serve it
@app.post("/predict")
def predict(lead: Lead):
record = normalize_record(lead.model_dump())
prob = float(pipeline.predict_proba([record])[0, 1])
return {"conversion_probability": round(prob, 6),
        "conversion": prob &amp;gt;= 0.5}
uv run uvicorn predict:app --host 0.0.0.0 --port 9696
curl -s &lt;a href="http://localhost:9696/health" rel="noopener noreferrer"&gt;http://localhost:9696/health&lt;/a&gt;
# {"status":"ok","model_sha256":"1646...","feature_defaults_sha256":"1694..."}
Note the API loads the pipeline once at startup. Swap the artifact, restart uvicorn.&lt;/li&gt;
&lt;li&gt;Containerize and confirm parity
FROM python:3.11.15-slim-bookworm&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;COPY --from=ghcr.io/astral-sh/uv:0.10.11 /uv /uvx /bin/&lt;/p&gt;

&lt;p&gt;ENV PATH="/code/.venv/bin:$PATH" \&lt;br&gt;
    UV_COMPILE_BYTECODE=1 \&lt;br&gt;
    UV_LINK_MODE=copy&lt;/p&gt;

&lt;p&gt;WORKDIR /code&lt;br&gt;
COPY pyproject.toml uv.lock .python-version ./&lt;br&gt;
RUN uv sync --locked --no-dev&lt;/p&gt;

&lt;p&gt;COPY model.py predict.py feature_defaults.json model_metadata.json pipeline.bin ./&lt;/p&gt;

&lt;p&gt;EXPOSE 9696&lt;br&gt;
ENTRYPOINT ["uvicorn", "predict:app", "--host", "0.0.0.0", "--port", "9696"]&lt;br&gt;
docker build -t zoomcamp-model:2026-hw5 .&lt;br&gt;
docker run --rm -p 9696:9696 zoomcamp-model:2026-hw5&lt;br&gt;
python q6_test.py&lt;/p&gt;

&lt;h1&gt;
  
  
  {'conversion_probability': 0.769799, 'conversion': True}
&lt;/h1&gt;

&lt;p&gt;Same inference locally and in the container. If yours differ, the build context has a different artifact or lockfile. Free port 9696 before running (stop local uvicorn first).&lt;/p&gt;

&lt;p&gt;What stuck with me&lt;br&gt;
Serialize the whole pipeline, not just the model, so preprocessing travels with it. And verify before you debug. Most of my "model bugs" were environment bugs.&lt;/p&gt;

&lt;p&gt;Full code, notebook and FAQ are here: &lt;a href="https://github.com/marcos-Nieves-24/machine-learning-zoomcamp" rel="noopener noreferrer"&gt;https://github.com/marcos-Nieves-24/machine-learning-zoomcamp&lt;/a&gt;&lt;/p&gt;

</description>
      <category>machinelearning</category>
      <category>ai</category>
      <category>zoomcamp</category>
      <category>programming</category>
    </item>
  </channel>
</rss>
