๐ฑ A Design & Development Company ๐ We love working with SMBs and enterprises ๐ค @ycombinator W14 company
San Francisco, CA
Joined July 2015
- Tweets671
- Following754
- Followers360
- Likes1.5K
Planning an identity provider migration?
Do you know how many of your users don't have a password?
Or which services still trust your old issuer?
These rarely show up in the provider's migration guide. Swipe for the three breakpoints we see most. ๐
Which one has caught you out before? Tell us in the comments.
The full checklist is coming soon. ๐
Since 2020, weโve worked across 40+ IoT products, helping companies solve technical challenges, develop new capabilities, and support products beyond launch.
We can take on the full build or work alongside your existing IoT team where additional expertise is needed.
Our new IoT expertise page brings together our experience, client references, and the areas where we can support your next product milestone.
Have an IoT product that needs some extra engineering expertise? Letโs talk ๐ povio.com/lets-talk
AI can write syntax faster than anyone can type it ๐งฉ
What it doesnโt decide is how the pieces fit together, or where the system breaks first under real load.
That still takes someone who understands the whole system ๐
Backend: developers who understand how the UI consumes data stop building endpoints nobody needed. Extra fields and redundant options only add maintenance cost.
The engineers shipping fastest right now move between both layers without losing time in handoffs.
Looking for engineers who can work across the whole stack? Letโs talk ๐povio.com/lets-talk
What does one finished AI task cost once retries are counted? ๐ธ
Most teams can quote the price of a model call. Few can say how many calls it takes to get one good answer.
All five cost drivers on one page ๐
An agent performs well in every demo ๐ฌ Then it ships, meets real data, and returns answers that are technically correct and completely useless ๐
The model didnโt change. The context around it did. A thread on context engineering ๐งต
1. ๐๐ต๐๐ป๐ธ๐ถ๐ป๐ด. How you split data before it reaches the model decides what it can retrieve later. Chunk it wrong and the model loses the connections between related pieces.
2. ๐ฅ๐ฒ๐๐ฟ๐ถ๐ฒ๐๐ฎ๐น ๐ณ๐ฟ๐ฎ๐บ๐ถ๐ป๐ด. Pulling the right data isnโt enough. How itโs framed for the model, whatโs labeled as relevant versus background, decides whether the model uses it correctly.
3. ๐ฆ๐ถ๐ด๐ป๐ฎ๐น ๐ณ๐ถ๐น๐๐ฒ๐ฟ๐ถ๐ป๐ด. More context isnโt always better. Feeding a model everything it might need often buries the one detail that mattered, so filtering out noise is part of the job.
Why teams miss it: traditional ML training covers models and data pipelines, not how information is structured for a model at inference time.
Building AI agents for production? Letโs talk ๐povio.com/lets-talk
AI can now handle a surprising amount of project management work. ๐๏ธ
It can pull context from old decisions, compare requests against the backlog, prepare briefs, surface dependencies, and draft updates.
There are still a few moments where we want a human PM in the middle:
5. When someone needs to own the outcome
AI can recommend. A PM still needs to stand behind the plan, communicate the trade-off, and take responsibility for what happens next.
The admin part of project management got easier. Status summaries, meeting notes, first drafts.
The hard part got harder. It's now about knowing what's been checked, by whom, and to what standard.
Four things that changed about the job ๐
A lot of project management time disappears into finding context, translating requests, and keeping decisions aligned.
David has built his AI workflow around reducing that overhead. Hereโs how: ๐งต
5. Review independently, then decide
Use multiple AI systems for additional checks while keeping consequential decisions under human oversight.
6. Feed the learning back in
Verified outcomes become part of the project knowledge for future work.