DEV Community

Mukunda Rao Katta
Mukunda Rao Katta

Posted on

What I shipped in a weekend: 3 Python libraries, 6 articles, a Homebrew formula, and an upstream PR

#showdev #python #opensource #productivity

Two days, end-to-end. Here is the log of what shipped, what blocked, and what the time actually went into.

The list

Three Python libraries on PyPI:

  • bedrockcache: static auditor for Anthropic prompt caching across six Bedrock backends.
  • bedrockstack: Bedrock-aware retry policy, cost ledger, and AnthropicBedrock streaming-error normalizer.
  • ragvitals: five-dimensional production drift detection for RAG (still GitHub-only; PyPI was rate-limited at the time of writing).

A companion demo repo with synthetic + real-models harnesses: ragvitals-gemma-demo.

One Homebrew formula:

brew install MukundaKatta/tap/bedrockcache
Enter fullscreen mode Exit fullscreen mode

Six articles on dev.to:

  1. Your RAG works on Claude. Does it work on Gemma 4?
  2. Fully open-source RAG with pgvector + pgai + Ollama
  3. Hardening a Notion MCP workflow with seven small utility MCP servers
  4. Agents that monitor themselves on Tiger's Agentic Postgres
  5. Hardening Algolia MCP Server the same way I hardened Notion MCP
  6. A 60-line Redis sink for ragvitals

One upstream PR, currently open: strands-agents/sdk-python#2301. Adds an optional s3_client= kwarg to S3SessionManager so callers can reuse a pre-built boto3 client instead of paying for boto3.Session(...) + session.client("s3", ...) on every manager instance. Maintainable, additive, backward-compatible. 50/50 session tests pass locally.

What actually consumed the time

If you guessed "writing code," you would be wrong. Here is where the hours actually went, sorted by surprise factor:

1. PyPI's "Too many new projects created" rate limit (lost about 8 hours)

I had three brand-new package names to register. PyPI has a per-account quota on first-time project creation that turns out to be tighter than the docs suggest. The rate limit is per-account, sliding-window, and shows up as 429 Too Many Requests with body Too many new projects created. It does not respect Retry-After because it is not a per-second rate limit; it is a quota.

bedrockcache and bedrockstack got through after staggered retries. ragvitals still has not, despite 8+ attempts spread across hours. The unblock is to email admin@pypi.org with the project names; the response time is typically <24h.

Lesson: for any future PyPI-first launch, register the project names with twine upload of a 0.0.0a0 placeholder weeks before the real release. Names + versions both have per-account limits, but creating a project for the first time is the limit that bit me.

2. Name squatting on PyPI (lost 30 minutes + a rename)

The package I originally called ragdrift was squatted on PyPI by an unrelated person who uploaded a 0.1.0 in March and never touched it again. PyPI's PEP 541 takeover process takes weeks. The cheap fix was to rename to ragvitals, which is arguably a better name anyway because drift is only one of the five dimensions the library tracks.

Lesson: check https://pypi.org/pypi/<name>/json returns 404 BEFORE picking a name. I now do this as the first step of every new library scaffold.

3. dev.to's front-matter precedence over API fields (lost 15 minutes)

The dev.to REST API takes a published: true field on POST /api/articles. If your body_markdown contains YAML front-matter with published: false, the front-matter wins. My first publish came back as HTTP 201, looked like it worked, but the public URL 404'd because the front-matter still said published: false.

Lesson: when posting via the API, either set published: true in both the front-matter and the API field, or strip the published line from the front-matter entirely.

4. Em-dash stripping (recurring, low-cost)

I use em dashes naturally when writing. My voice rule excludes them from public posts. Every article needs a pre-publish grep -c "—" pass to confirm zero. Caught 3-5 stragglers per piece. Took 2 minutes per article; cost-of-living tax.

5. Cross-stack honesty (the real work)

The bulk of the actual hours went into making sure each article's claims were defensible:

  • The Gemma 4 article reports ResponseQuality.faithfulness=0.7858 when you swap from Claude to Gemma 4 in a synthetic harness. That exact number is pinned by a pytest test in the companion repo so CI catches it if the math drifts.
  • The pgai + Ollama article describes a demo/pgai_ollama_run.py script that actually exists in the repo, with the SQL it shows in the prose.
  • The Notion MCP hardening recipe references seven existing @mukundakatta MCP servers I had already shipped. None of them are vapor; the article's value is the composition, not the components.
  • The Tiger Agentic Postgres article is the one place I explicitly flagged a non-runnable piece (the drift_report.sql is inline in the post, not committed to the repo yet). Naming the gap in the post itself is the only honest move.

The token-and-paste-velocity that makes "ship 6 articles in 2 days" possible is exactly what makes it tempting to fabricate one number, one repo link, one phantom function. Forcing each piece to have a runnable artifact in a public repo is the only thing that kept the velocity from outpacing the truth.

What did not work

  • The Notion MCP article assumes the seven utility servers are individually well-tested. They are CI-green at the package level, but I have not personally run the full 7-MCP stack against a live Notion workspace. The article's recipe is correct; my evidence is the component-level tests, not an end-to-end demo.
  • Amazon Q Developer Quack The Code Challenge: skipped. The theme was specifically about Amazon Q Developer CLI usage. I have not actually used Q Developer in production for these libraries. Writing a fictional "I used Q to build bedrockcache" piece would have been dishonest, so I dropped it from the batch.
  • Trusted Publishing setup: deferred. Once the PyPI quota clears, I want every future version to publish from a git tag v0.x.y && git push --tags via GitHub Actions, no token in chat. The workflow YAML is straightforward; the bottleneck is that PyPI's "pending publishers" UI requires a browser click I cannot script.

What I'd do differently

  1. Pre-register PyPI names with a placeholder version the moment I commit to the name. Eliminates the new-project rate-limit class of failure entirely.
  2. Write the article and the test that pins its numbers in the same commit. I did this for the Gemma 4 piece but not for the others; the discipline made the Gemma article the strongest of the six.
  3. Stage the dev.to publishes 30 minutes apart, not back-to-back. dev.to does not rate-limit posts the way PyPI rate-limits projects, but five articles in two hours feels spammy to followers. Cadence is its own form of quality.
  4. One repo for the demo, six article markdowns inside it. I did this; recommend it. The article markdowns in the repo (article-*.md) double as draft history AND act as canonical source for re-publishing on other surfaces.

What I'd do next

The libraries point at a wider gap I want to close: a single open-source "AWS Bedrock + Claude in production" toolkit that bundles bedrockcache + bedrockstack + ragvitals behind a coherent CLI and a documented pipeline pattern. Today they are three sibling libraries with parallel READMEs. A 200-page tutorial repo with a cookiecutter-bedrock-rag template would land them in users' projects with one command instead of three pip installs and a config decision.

If you are running Bedrock + Anthropic in production and any of these libraries would have caught a bug you actually hit, comment with what the bug was. The boring stuff (a $38k prompt-caching miss, a streaming-error parity gap, a re-indexed corpus that silently lost recall) is exactly the kind of signal worth feeding back into v0.2.

Repos referenced

Two days, eleven public artifacts. Tools that turned the velocity from "tweet-thread" into "shippable" were uv, hatchling, the dev.to REST API, and a grep -c "—" line in my pre-publish checklist.

More in this series

Top comments (0)