product engineer | exploring the world beyond software. into robots. i build things
Frankfurt
Joined January 2017
- Tweets2.4K
- Following581
- Followers286
- Likes777
Vercel took its lead qualification team from 10 people to 1. One person now runs the US, and about 20% of another covers Europe and APAC.
Every agent is built by three people in a tripod:
- go-to-market engineer
- data scientist
- a specialist for the domain
For lead qualification, the engineer shadowed one of their sales development reps for a few days. Each step and each opened browser tab were recorded (LinkedIn, CRM, Slack). "We documented it, made the workflow legible before touching any AI."
One weekend and the first version was built.
Here's the part many skip: the agent ran in shadow mode for six weeks, and the same SDR reviewed every single output and gave feedback until she "couldn't improve it anymore."
The result:
- agent costs under $5,000/yr in infra and tokens, plus 20% of one engineer's time to maintain
- it performs like a 90th-percentile rep 100% of the time, 24/7
- other reps moved to higher-value roles
- same method was used on 30 other SDR workflows and raised SDR quotas by 30% this quarter
A few other rules:
1) A shared data layer sits under every agent, mapping what each business term means in the data. "Without it, agents hallucinate or give generic answers."
2) Their meeting agent lives in Slack with no UI, and only works because other tools have APIs and webhooks (e.g. Salesforce). "If they didn't, we would have gotten rid of the tool."
3) Their support agent handles 93% of cases. Three engineers work on it. They built it in two months, after an off-the-shelf tool didn't give enough results.
4) Every quarter, she asks if another company has built something better, faster or cheaper. If yes, "our agent's going to go."
Notice how little of this is about better models. Most of the effort went into understanding how work actually gets done, making it legible, and only then automating it.
"Hardware engineers have baby software, built in spreadsheets..."
12 years ago I founded Boom, as a software engineer doing hardware for the first time. I’m probably the least handy founder in hard tech. My 12 y/o son does my soldering and 3D printing for home projects, and I’ve struggled with imposter syndrome the whole way through. What business does a software guy have building supersonic jets and turbines?
More than a decade in, the most insane realization is that software is the key to reaccelerating hardware engineering. But in a way that I don’t think is widely understood or implemented, even in the burgeoning hard tech community.
Most hardware engineering is actually software engineering in disguise. It’s spreadsheet “code,” engineering calculations done as one-offs. Eg wing structural loads in, spar sizing design out. Thermal loads in, air conditioning system sizing out.
Everything is a calculation, and calculation is software. But these calculations are done essentially by hand, as if we are doing arithmetic but have not yet discovered algebra.
Hardware engineers have baby software, built in spreadsheets, the kind of crappy software I wrote as a kid before I understood architecture, unit tests, APIs, version control, and everything else that makes modern software good.
Our airplanes and engines are built with the software sophistication of a script kiddie.
So what we tell our engineers: don’t design the hardware, design the *algorithm* for designing the hardware. And take the algorithm seriously—as an algorithm and as a first-class engineered software system.
Never design a piece of hardware, such that you couldn’t redesign it tomorrow with a different set of assumptions.
This all sounds obvious, yet our experience at Boom is that it changes everything.
Ten years ago we couldn’t afford to build much software. We had a small team—and there were few software engineers with the curiosity, brilliance, and vision required to learn hardware engineering.
But now AI changes the equation. Now everyone can code. The software engineers define the architecture. The hardware people bring the physics and engineering know-how.
We have a turbomachinery design tool that lets one engineer do in realtime what previously took 100 engineers six months. Another tool auto-generates 65% of entire aircraft structure. I’m sure we’ll figure out the other 35% too. Eventually these should not be multiple tools, but one integrated physics model of a complete vehicle, airframe and powertrain.
The results speak for themselves. A boomless supersonic airplane built by 50. A full scale supersonic turbine made by 150. And with this approach, I don’t think we’ll need that many more people to build and certify a supersonic airliner.
You don’t need a big team to build big things—not when everyone on the team is a software engineer.
I almost scrolled past this tweet.
I wasn't even aware of what was happening on the market yesterday 😅
So I did what I always do in those moments: I bought the dip.
Let's see, but today it's all ↗️
Generational moment to buy...
Why? Let me explain, in under 60 seconds, the game on the field for Team Blue:
First, the democrats aren't as radical as the GOP painted them (just like MAGA isn't as bad as Dems painted them)
Second, the Democrats can't win long term if they kill the markets (aka the AI trade) with regulations, taxes and by vilifying the job creators.
It is critical that we support moderates in the Democratic Party like @RahmEmanuel and steer the rest of Team Blue toward being pro-capitalism, AKA pro-jobs and economic growth.
The last thing democrats want to be associated with is tanking people's 401ks and causing layoffs and a recession.
@RepJeffries understands this...
I loved Ben's bookshelf idea, so I copied it for my own site.
I’ve tracked my reading for ten years, first in a spreadsheet, then in Notion. Both felt like a database though.
The cool part is the data costs nothing:
- Open Library gives the cover and page count for free, with no API key
- A 90-page book gets a 26px spine, a 700-page one gets 72px
- Spine color comes from the cover image
- Heights vary a little based on a hash of the book's slug, so the tops aren't flat
The only cost, which is optional, is for realistic book spines. Mine are pure css now.
I tried that once. Paid ca. $5 for 140 books ($0.03/call via treg .to's reAPI). But images came out at the wrong scale. Before the next run, I need to fix the aspect ratio and the prompt passed to Gemini. Only later did I notice that the first prompt said: no illustration, no photograph, no texture, no pattern, no gradient.
Here’s how it works:
1. The data lives in a Notion database, with one row per book: title, author and the year I read it
2. A script pulls the rows through the Notion API
3. For each book, it asks Open Library for the cover and the page count. It's free, no API key
4. Covers are downloaded and served from my own domain, not hotlinked. My site’s security policy (CSP) only allows images from itself. A bonus: the shelf still works when Open Library is slow
5. Spine width follows the page count. A 90-page book is 26px wide, a 700-page book is 72px
6. Spine colour is taken from the cover image (for the css option)
7. Heights vary a little, based on a hash of each book’s slug. So the tops aren’t flat like a bar chart, but they stay the same on every build.
I added a /ramble command to my Hermes agent because I prefer to think out loud.
The first sentences are messy, mostly rambling, and full of uhms and other fillers. But after a while, a structure shows up, and the agent turns that text into notes for me.
It mines for ideas, insights and angles for a post. The agent also says which ideas were dropped and why.
For dictating, I use Natter. I could probably also use voice memos, but with Natter I can see the text, which helps me pause, look back at what I have said and then continue.
Oh, and I use the agent as a Telegram bot.
The holding company model sounds attractive. But buying and fixing an existing business might actually be harder than starting one.
I see this when I scout companies for acquisitions. Many have a sound business model, good financials, happy & paying customers - they're good businesses.
But in smaller comps, the owner IS the business. He talks to customers, sets prices & discounts (often in his head), handles supplier deals. All of the rules are in one person’s head.
Throwing AI at it won't help, you can't automate what nobody wrote down. And buyers are hesitant to acquire a business that leaves with the owner.
In a startup, nobody has to unlearn anything, there's no muscle memory. In an established company, with customers, and cash flow, every process has a reason. Someone has fought for its existence. People will defend it.
That's not to say it's a bad model, no. Many are doing this, and it can be lucrative.
What I'm saying is that it's challenging for different reasons (see above). People considering this need a different skill set. It's not as simple as prompting your agent.
ai makes 1 to n dramatically easier, but it doesn't make 0 to 1 proportionally easier. the latter requires all of the things that were required pre ai whereas with the former, you can ship ideas extremely quickly, sometimes within hours. this is not to say ai can't help you achieve pmf, obviously it does. i am just saying that it is not proportional.
what's also a new challenge with zero to one now is that whatever you do may also be simply copied very quickly with ai either by incumbents & competitors.
this is why there are now holding companies acquiring existing businesses at scale & transforming them to remove the countervailing force of org bureaucracy &/or lack of structural capability to leverage ai to transform the business.
Last week I finally switched from Ghostty to Claude Desktop and I'm not going back.
I stayed in the terminal mostly out of habit. But I also didn't like the initial version of the desktop app. But small things in the terminal were annoying like copying text or looking through sessions.
It took me a couple of days to set it up my way and get used to it.
Now I don’t even open Cursor IDE.
a company with $10M+ in revenue hired 3 college students to act as its AI department.
the problem was overhead.
the company grew fast, and its costs grew just as fast. then growth stopped but costs didn't.
so the CEO wanted to know: how do we keep growing without hiring more people?
↓
5/
that one demo turned into a map of 19 workflows. now they build them and train the staff.
(note: these are the founder's own numbers, on a show that sells a course on this exact model. the company isn't named, and there are no results yet)
That's a massive step towards fighting scammers.
Generated images are everywhere, and catching them is hard. Ordinary people get scammed on many platforms, like dating and real estate.
Pangram claims 99.9% accuracy on images and video. The API and Chrome extension are live today.
China generated about 2.4 times as much electricity as the US in 2025.
It added over 400 GW of wind and solar.
The AI race comes down to who has spare power, not who has the best chips.
I think the big money in AI is moving from building models to deploying them.
and consultancies will own this work because they already own customer relationships.
no, consulting is not dead
Mistral is doubling down on its plan to become a boutique AI consultancy.
In 2025, they had 132 engineers working inside customer companies. By 2029, they plan to have 2,000. Their careers page lists far more roles than research jobs.
Some people say Mistral is turning into "Capgemini with GPUs".
I read it differently.
It’s closer to Palantir: put engineers next to the customer, learn the messy process, then turn what you learn into product.
For Europe this might actually be a more realistic way.
A German manufacturer or a French bank doesn’t need the biggest model but rather someone who can make a model work for their particular use cases.
My only concern is about margins.
Consultants earn per hour. Model companies earn per token. That's a different business model and more challenging to scale.
There are 2 ways to build a data center in orbit:
Pay $75–100M, like Starcloud was told. Or pay about $2.5M, like they say they did.
More facts you should know:
> raised $450M, valued at about $2.3B
> Starcloud 1 went up last November with the first Nvidia H100 in space, and trained a small model (Karpathy's nanoGPT) up there
> they were told it would cost $75–100M. they say they built it for about $2.5M
> they bought a broken, secondhand vibration table for $200k. a new one costs $1.5M and takes almost 2 years to arrive
> Starcloud 2 launches in January with 100x the power of the first one
> Starcloud 3 weighs about 2 tons, and 50 of them fit in one Starship
> they've filed with the FCC for 88,000 of them
> the cash in the bank earns about $1M a week in interest
The whole plan rests on one number:
Space beats Earth once a kilo to orbit costs less than about $500, the CEO says.
On Falcon 9 it's about $5,000 today.