Company Product Builder. Co-Founder/ex-CTO ACME Ticketing. Now focus on native AI apps. Hands -on, founder led mode .
Redwood City, CA
Joined December 2007
- Tweets2.5K
- Following1.2K
- Followers536
- Likes2.8K
Pinned Tweet
Actionable framework to deal with tech debt as your SaaS company matures. Hard-earned wisdom.
linkedin.com/pulse/how-deal-…
Re-analyzing my code to see where @typesafeai can cause an accuracy lift . In small sections so far, yet in high volume agentic tool calls. let's see where it lands eval wise, at least for baseline, then cost gains.
💯 & also my point in podcasts.apple.com/us/podcas…
@shaneparrish would love to connect.
"The criticism that a lot of people direct at companies is that companies are short-term focused,"
But why? @tobi answers
'I don't think the executives tend to be short-term focused, but they are incentivize to keep their job.'
"Therefore, they need to be able to prove that they're doing a good job at intervals."
that means a lot of companies are in trouble when major change happens.
"The perfect thing for a company to do is rebuild their entire product from the ground up for the AI age, which is gonna take a while. [But] They won't do it because the short-term incentive is not there."
"Their local incentive system is quarterly attaboys."
As Munger said, "Show me the incentives, and I'll show you the outcome."
Decided to switch off @twilio . Horrible doc not up to date, bad customer support on lower plans. Not a way to grow revenue down the line from upgrades . 20y ago, i loved Twillio, don't recognize it today.
Friends, droped podcast #1 on a series of AI related topics. This one is about how to better transform to AI native products. Feedback welcome #transformAI
podcasts.apple.com/us/podcas…
When you code, put your institutional knowledge , mostly best practices, in the Claude.md files ( or ~) and you are a step closer to self-learning codebase changes. Will share soon! #AIdev
Automating your app dev as much as possible. I enjoy ideating to code quickly, therefore i set up for automation upfront, almost manual self learn the env if you will, see below ##ArtificialIntelligence :
5/ big time use of Github workflows to automate CI/CD to fully regress , including evals
6/ api led architecture lead to mcp set up and then CLIs, makes it easy to create all kind of tools to test the app
A big fan of custom models in SML. They work very well, especially when you tie inference prompt/schemas to what it was trained with. Big qualify lift. Thanks @tobi for sharing.
Btw we open sourced the core infra piece that makes these self improving loops possible. tangleml.com
1/ API boundaries
2/ shape your tools for mcp around workflows
3/ Headless from the get go
4/ graph your agents & re-inforce with prompts for more reliability
4/ CLIs for agents to easily consume your functions
5 /Include classical encapsulation pattern (LLM context centralized, services under APIs ..)
6/ Heavily use unstructured document like DB , they work really well with context/conversations
7/ Scope your memory arch early on to retain past knowledge
8/ eval where non deterministic
That will get a solid architecture out of the box to iterate on quickly and adding features at a blistering pace. Much of the above can be tooled to enforce it all , then speed develops. #coding
In a world where AI can code so much, your architectural pattern boundaries is what makes the code maintainable, re-usable. #AI #buildinpublic #AgenticAI
sundays = coding besides night hacking. Incredible how SaaS app can be rewritten nowadays into new native AI stacks. For incumbents, good to go back to Day 1 thinking & founder led execution mode.
A big fan of @stripe's acq. of @OpenRouter. Pricing learnings from card network variables costs transfer to tokens naturally. Most new AI like companies now need to package pricing into agentic tasks and learn how to price their outcome against their token costs.
Brilliant move.
The days of a one person startup are firmly here ( code is AI based mostly, web sites can do sales with co-pilots). Obviously, to scale you need more folks esp if larger customers , yet 2-3 folks can grow fast nowadays, even in b2b. wdyt ? cc @ycombinator
A big fan of @deepseek_ai ! can't wait to try latest and see evals lift.
We’re launching DeepSeek-V4-Pro today! 🚀
🔷 Major Agent upgrades with strong production gains!
🔷 Flexible reasoning effort for V4-Pro & V4-Flash: low for simple tasks, high for daily Agent workflows, max for complex tasks.
🔷 Native OpenAI Responses API support, optimized for Codex with one-click setup.
V4 Pro is now available on app/web. Try it via “Expert Mode”.
V4 Pro is also available via API. Model names remain unchanged—please refer to the API docs for setup details.
