
Bootstrapped, fast-growing, profitable. Founder & CEO @BRIDGE_IN_COM, father of 3. Anti-bureaucracy tech optimist building proof Europe can still win.
Lisbon, Portugal
Joined March 2010
- Tweets3.5K
- Following853
- Followers506
- Likes373
We spent ten years moving everything to the cloud. #AI is making some of us pull some of it back.
Every prompt we run through someone else's model sends data past a perimeter we do not control. For a law firm, a compliance company, anyone whose product is trust, that adds up.
The useful way to think about it is
four separate questions, building on the idea of data sovereignty.
- Does the model stay the same, or can it change without warning?
- Does a third party see how we use AI?
- While the model runs, can the operator read the machine?
- Whose law can reach the data?
Run every hosting option through those four. A hosted API loses all of
them. A rented open-weight model on someone's cloud gives back two.
Confidential computing closes a third. Only on-premises wins all four, but you'll need to pay for it in maintenance like in the old days.
The line that surprises people most: jurisdiction follows the operator,
not the server. A US provider sits under the CLOUD Act even with the
servers in Frankfurt. European "sovereign cloud" is often still US law,
one hop away.
We probably won't go back to racks in closets. But the default is no longer
automatic. Companies are starting to choose, deliberately, who gets to
touch their intelligence.
2
2
62
Dear @AnthropicAI why the hell is it so hard to pay you? Your authorization and collection with a distinct merchant make all my banks block you for suspected fraud. Every single month, the same! And you block access to @claudeai immediately (which is predatory). Can you please have a funds collection mechanism that doesn't suck?
For one of the most valuable IPOs in the world, with the gods of intelligence on your side, you are unbelievably bad at the basic feature of letting your customers pay you.
29
Do you have a fleet of AI agents?
I've been spending a lot of time building ours with Hermes from @NousResearch.
- Not an engineering software coding harness. We already have that for 1 year, and it's easier because users are technical and evals are easier.
- Not customer-facing agents. We are in a highly regulated area, and our certified humans carry the responsibility. Some solutions are popping up, but no AI agent is a certified lawyer or chartered accountant.
- We are building a business automation AI fleet where every human has a side-helper. This has been genuinely hard because it needs to serve non-technical users, it needs to evolve and adapt to its supervisor, but it needs to follow BRIDGE IN's constitution and security stance, plus it needs to evolve fleet capabilities and skills through crowd wisdom.
Jensen (CEO at Nvidia), a big proponent of open-source and not a doomer with a god-like complex, closed a recent keynote with a Hermes Agent demo saying, "the way to build an agent is to ask an agent to build it for you".
Which is what we are doing, although with some frustration on shared learning and fleet coordination. Still, I believe we are betting on the right horse.
I will say that Nous Research motto "Rebellion to tyrants is obedience to God", which comes from Benjamin Franklin (US founding father) proposal to the Great Seal of the United States, depicting the moral imperative of fighting oppressive, unjust rulers or governments, never felt more omniscient, in a world of polarized politics and AI labs thinking they are neo-gods.
2
2
152
In #Hermes, currently, I can't pick a different model per delegating child. The child gets the session model, or a global override that pins all children.
This is already being asked: there's an open mergeable PR (#107945) with months of community discussion and zero maintainer review. Nine-plus PRs sit on the same gap.
A genuinely useful agent harness lets a skill decide HOW to run a subtask: cheap model, vision model, private model. We're builders on your harness. Can we get the per-task model param? @NousResearch
44
#AI agent plasticity paper from Meta Superintelligence Labs
The three surprising findings:
1. The best player isn't the best learner. The agent that ends up with the highest score is often a different one from the agent that improves the fastest per dollar. Being talented and being coachable are two different things.
2. Some agents just don't get better. In chess, Go and Hex, one model (Claude Fable 5) ends up best, but another (GPT-5.6 Sol) learns the most per dollar. In NetHack, only one model (Claude Opus 5.5) improved meaningfully, and it cost about $1,073 of "learning" to gain 66 points. The others basically stayed flat or got worse.
3. The reason slow learners stay slow is embarrassing: they write notes and then never read them. The fast learners actually reuse their own notes, but even they fail when the note they wrote is bad quality. So the bottleneck is either "you don't reuse your own work" or "your work isn't good enough to reuse."
Paper: arxiv.org/abs/2610.08902
78
A lot of the tech you use every day is priced below what it actually costs to run.
We are hearing that about AI tokens, but this is an old playbook.
In software that's easy to hide.
One more subscriber costs almost nothing to serve, so a company can lose money on every customer for years and still call the chart "growth" for the next fundraising.
The playbook is well known. Undercut everyone, buy the market, worry about profit later, once the competitors who couldn't burn cash that long are gone.
Uber ran that playbook for a decade and still isn't the last one standing.
Recently, it agreed to buy Delivery Hero for around $15B, a deal that would put it in close to 100 countries and make it one of the biggest delivery platforms outside China.
That's what the endgame tends to look like. The market keeps concentrating until a handful of players are the ones setting the terms.
The people applauding this rarely ask what funds it. Usually, it's someone else's capital. Round after round covers the losses, keeps the discounts alive, and props up a business with unit economics that don't work yet.
From the outside, it looks like momentum. It's really a company falling forward, betting it turns into flight.
When the funding stops, and eventually it does, the fall is quick.
Teams hired against projections get cut.
Suppliers who built around one big client are left exposed.
For every founder who exits rich, there's a list of people who stopped getting paid so someone else's story could end well.
The customer feels like the winner through all of this.
Cheap now, cheaper next year.
What's hard to see is that a price held below cost is doing a job. It's buying your dependence. Once you depend on it, the price stops being yours to negotiate.
I built @BRIDGE_IN_COM without #VC money.
We were profitable about 18 months in.
When there's no outside money funding the discount, the math is simple.
The customer pays what the service is worth, and that payment is what keeps the company alive. No burn rate to defend, so nothing to protect except the client actually paying you.
Anti-dumping rules already exist, but enforcement is loose in tech.
We should let startups find PMF and do their price discovery. Losing money to learn a market is not the same as building a business that only survives while someone keeps funding the discount.
Anti-price-dumping measures should be enforced once a company reaches real scale, to help customers tell the difference between a fair price and one that's subsidised to keep them dependent.
1
72
We are now in a world where the first reaction to unexpected new information is: this is probably fake.
Social media already spreads false stories, but with #AI, anyone can generate a video of a head of state saying something they never said, in seconds.
We now have trust issues.
1
2
29
This is what I keep telling our team at @BRIDGE_IN_COM: customers don't buy processes or software. They buy trust.
Trust that we keep them compliant and safe. Trust that certified professionals put their name to the work.
No #AI model does that part.
Do you know if you are paying for unused Claude subscriptions?
We were.
This is about the soft dark pattern I found while trying to address it.
- Usage is hard to see. No clean per-user last-active view, unlike most SaaS. Finding idle paid users means stitching exports across Chat, Code, and Cowork. This was strangely hard for a company that otherwise makes available a lot of useful analytics.
- Removing a member does not stop the payment. The seat stays paid until you separately reduce the allocation. This is not clearly disclosed, and I didn't expect it. It is the practice of some bigger players like Microsoft. Splitting it into two actions adds friction, and that seems on purpose.
- License proration only runs one way. Adding a user bills immediately, in one step, with no separate seat action. Reducing waits for the end of the period, with no matching fast credit. There is a deliberate asymmetry to benefit Anthropic.
- Seat reduction is clumsy and wasteful. Remove a user and you keep paying for that seat until the end of the billing period, so you pay for a seat nobody can use, as you can't schedule a seat reduction without removing the user. To change the reduction after scheduling it, you need to cancel the pending downgrade and redo it. Downgrades also ask for an extra confirmation. This is too much evidence to be an accident. It is deliberately created friction, soft enough not to amount to the worst dark patterns in the industry, but still. Sneaky.
Credit where it is due: we are paying for Claude on top of Gemini because it is good.
But I dislike the deceptive posture.
48
Bureaucracy in EU is bad and often non-logical. In PT banks somehow manage to be worse than government.
This frustrated me so much I started @BRIDGE_IN_COM to handle it.
104
We don't sell accounting services.
We provide peace of mind to business owners.
@BRIDGE_IN_COM doesn't sell accounting services, payroll, EOR, or legal support.
Our mission is to make the bureaucracy of running a company easy to comply with. And this means that we do provide these services and a lot more in between that are often left to no one.
Nobody wants to pay for accounting.
People want to be able to run their business without issues or fines from the tax authorities. The less trouble business owners have chasing down that missing paper invoice or that last-minute filing, the better.
Less time spent with bureaucracy means more time growing the business.
We are running a limited program for startups in #Portugal.
Get in touch if you want to stop cursing at bureaucracy.
1
1
34
Strip out inflation, and the EOR market is growing about 4% a year.
Now ask why anyone raised billions to win that market, and what happens when that math catches up.
Let's dig in.
Line up the EOR shortlist you were comparing in 2022. Four of those names have since been acquired, renamed, or quietly put up for sale.
- Safeguard Global's payroll division was folded into Deel last year.
- Velocity Global, a name tied to this category for a decade, is now Pebl.
- Horizons is now Remote People.
- Papaya Global, worth 3.7 billion dollars at its peak, was reportedly in talks earlier this year to sell itself.
That is four names off a single shortlist, and not one of them sits where it did.
A simple analysis suggests this is just a market maturing, with fresh entrants replacing the old guard. Maybe so. But look at the growth underneath the churn. EOR is expanding around 7% a year. Eurozone inflation is running near 3%, so in real terms the category is compounding closer to 4%. That is an ordinary services business, not the kind of curve that justifies the valuations and the leverage these companies raised on a promise of software-style hypergrowth.
When the capital going in needs a 10x return and the market it sits on grows 4% after inflation, the gap has to close somewhere. Usually it closes as an acquisition, or a down round, or a rebrand hiding a pivot or repositioning.
The rebrand deserves the most suspicion and can mask other issues. This is an old move, and outsourcing has run it for decades. Poor delivery, payroll mistakes, missing filings, support goes quiet, and the complaints collect in public. Glassdoor, Trustpilot, and others make this highly visible. A rename resets the search results, but the poor service quality and sometimes ethical shortcuts are still there. Same entities, same problems, but a new name.
Working with one of the big ones does not buy you calm either. Deel, possibly the largest player in the space, has spent more than a year defending a corporate espionage suit from Rippling, one a federal judge allowed to move ahead this year on racketeering grounds, with the US Justice Department reportedly taking an interest. Set aside who wins. A company pouring its attention into racketeering discovery is not a company giving full attention to your payroll.
BRIDGE IN has carried one name since 2020. There is no venture money on our cap table waiting for its 10x, no down round to bury, no rebrand to outrun. We built on certified professionals and deep compliance knowledge, turned profitable early, and we still trade under the name we started with.
So if a provider you depend on was acquired, renamed, or sued this year, ask a plain question before you renew: what actually changed, and does my contract still sit with the company I chose?
Can @BRIDGE_IN_COM help? We cover Portugal, Spain, Italy and Poland. Not the whole world, not yet. We would rather be the best option in four countries than a logo in a hundred.
61
Portugal has the most demanding software-certification regime in the world.
Claude thinks so.
At least in the context of billing software.
The world is a big place, and while I don't know all billing software requirements across 195 different countries, I can attest that Portugal has several strict rules indeed.
At @BRIDGE_IN_COM, we helped many tech companies and scaleups, several from the US and UK, to incorporate in Portugal. And sure enough, at some point we need to explain to their CFO that no, they could not use Xero or QuickBooks in Portugal. They needed to use one of the local, and sometimes clunky, billing software and ERP systems.
This is probably not the worst barrier to business, but it is still remarkable to notice that a smaller country has requirements stricter than bigger economies.
The rest of Europe is moving in the same direction. The European Union's VAT in the Digital Age (ViDA) requires structured e-invoicing and real-time digital reporting for cross-border B2B from July 2030, with domestic systems converging by 2035. The problem is, right now, every European country has different rules and therefore different billing software.
The era of using a single, generic version of #Xero or #QuickBooks to invoice customers globally is ending.
32
Payroll is arithmetic. One correct number per person per month, written in law.
A deterministic engine runs 1,000 payslips in under a microsecond and gets it right 100% of the time.
A frontier #LLM takes 10–30 minutes, consumes enough energy to charge 100 smartphones, and will still be wrong ~12 times a year.
Using #AI to compute payroll is not only less accurate. It is wildly inefficient.









