Unsure of which subsystem they are leveraging, likely Keycloak. Betting they have it replicated in a Kubernetes cluster against @awscloud.bsky.social RDS multi region and multi A-Z; but with slow queries + a heavily used DB that needs constant maintenance and vacuuming.
@anthropic.com suffering disruptions lately purely due to the login subsystem bottleneck; which I dare to state is not optimal, especially when US east wakes up, due to not having enough dedicated pods running + slow queries at tokens minting time in one of the flows of OIDC
Indie Hacking - March 2026 👋
🚀 Work: 168h building podhoc.com - Released!
😍 The Vibe: Productive and progressing!
🧐 The MRR: $0
GO Artemis GO
If you keep piling stuff up to your Claude code agent with interrumptions, it ends up forgetting all the requests or losing some so you must re-check them all at the end to ensure everything was covered
Any good collaboratively built documentation on good parenting?
Am I the only one that I got Docker Desktop broken as a result of upgrading mac OS to version 26.4?
My costume for the 2027 Carnival - Dario Amodei
I just released a new version of crosswire.online supporting internationalization and translations + a better 3D globe exploring experience!
#saas #indiehacking #hacking #product #news #outlets #crosswire #crosswireonline
I just rebuilt my personal website from scratch and updated it to contain all the information: davidpelayo.com
If the Telegram MCP skill for Claude code isn't working in a way that the Telegram bot does not respond to you: On a Mac, at least, is because you needed to run `brew install oven-sh/bun/bun` first
@awscloud.bsky.social is there a fast lane to apply for free credits on new SaaS products that we entrepreneurs create? Thanks in advance!
Solo founder developing it fully with @awscloud.bsky.social , TF, @github.com, @developers.google.com , and many more. Over 600 hours of developing. Now cooking the iOS and Android apps.
Pending approval of Google OAuth for the YouTube integration, and Firefox and Chrome web extensions. A journey!
This week I released my first product: Multiple sources into a podcast with AI? Up to 74 languages, 6 styles + your custom focus, target hours of length, weight how much each source should appear, voices and more, ultra fast an easy, at podhoc.com. Doubts? Reach out!
My new tech debt. are untested new features now
Opus died - switched to Sonnet. Cleaned up all stale branches. For each open PR, performing a full in-depth technical review finding issues by criticality + GO/NO GO decision. Then fix the findings and review again. Post on each PR.
Parallelism. Productivity 🚀
🪦 In-memoriam of traditional manual software engineering 🪦
I find that two main reasons for which you'd typically drop a side project or simply even not start it was due to lack of time and errors solving frustration (w/ time consumption).
We're now in the time when you can simply create most of the software you imagine in record time!
What's the name of the God you believe in? Opus.
So it took me around 3 months of developing the first live prototype: podhoc.com (@podhoc.bsky.social)
Generate audio capsules with AI of almost any content you can imagine, ultra fast and easy. Supporting 50+ languages, audio styles, or content distillation techniques.
Check it out!
Recommendation: If you write a feature with the AI model A - then review the PR with the AI model B (never the same models from the same vendors).
AI is nowhere near its "theoretical" potential.
Anthropic created a new metric called "observed exposure"; tracking what AI could do vs. what it is actually doing in the real world.
The gap is HUGE. We are barely scratching the surface.
Telling a Cursor agent to inspect and retrieve changes from the history of changes of another Cursor agent, the cheapest way, for which Antigravity with Gemini 3.1 Pro tells me where the actual content of the Cursor agent history is. It's 🤯
5) Starting to work on an implementation in which I'll introduce open source models in the loop via AWS SageMaker or Google Vertex to reduce costs by creating on-spot instances of the async job I want to perform (but this is WIP right now).
3) Use super cheap models with specific prompts when requesting text output - in case this is a task very much demanded.
4) Look for a balance between different providers to resolve an issue for which I do not have much knowledge about (sometimes "auto" in Cursor, sometimes Sonnet, sometimes Gemini)
It's hard to say - I do not own too much expertise on the matter.
But what I've been trying to do is: 1) Avoiding models jumping to migrate contexts (Claude code); 2) Be extremely specific about what is needed to solve an issue with expensive models like Opus [...]
It all comes down to how efficient you handle context tokens and what are you requesting to the models.
After a few months building up a SaaS that I am about to release, is hard to believe that part of the core functionality relies on how well these two are managed.
Indie Hacking - February 2026 👋
🚀 Work: 110h building the AI product (close to release!)
🏊♂️ The Grind: 20h of sports!
🥰 The Vibe: Productive and strong!
👌 The MRR: $0 (for now!)
🖼️ The Pyrenees
