applied ai @happyrobot / prev. co-founder & cto @gizatechxyz / prev. scaling prod ml infra @adidas @bbva / tech+optimism
Madrid
Joined August 2011
- Tweets925
- Following1.5K
- Followers2.3K
- Likes4.9K
Pinned Tweet
After 3 years building @gizatechxyz from scratch, I’m stepping down from my full-time role as co-founder & CTO. I’ll stay close through a short transition and keep cheering the team on and supporting where my input is needed.
We started at the frontier of ZKML (Zero-Knowledge Machine Learning), making model outputs provable. The market was early, protocols weren’t ready to embed models in their smart contracts. So we focused where we could help most, DeFi AI agents that automate the manual work and improve crypto UX.
That’s how Arma was born: an autonomous yield-optimization agent for stablecoins. It continuously scans venues, prices risk, and reallocates within explicit guardrails (liquidity, tvl, protocol risk etc.). Decisions are transparent, actions auditable, and the system is built to self-recover.
Along the way, we were privileged enough to be backed by top tier investors like @coinfund and @cbventures. We grew a hands-on engineering team to 13, set a high delivery/quality bar, and shipped the best foundational infrastructure for agentic finance that is in production today.
Today, Arma manages $30M+ in assets with 5k+ daily active users. None of this happens without the people who built it, thank you to every teammate who shipped through the many long nights.
Now Giza enters a new chapter. I’m biased but I’m very confident in the team behind it and I’ll be around to help where useful.
To our investors, partners, and the community that believed from day zero: thank you for the ride.
Fran Algaba retweeted
How do always-on, proactive agents like @Bot work?
Watch my lecture at Stanford CS146S:
2:35 How we got here
6:13 What changed in the models
10:06 Inside an always-on agent
18:07 The harness
27:26 Context engineering
34:26 Where this is going
implemented this in our internal rag and got a solid 12% perf boost in our evals by just using jev as the reranker. also quite solid in terms of costs!
this is where things are going. in order to start the recursive flywheel there needs to be a proper trace mining system which is 80% of the problem.
Fran Algaba retweeted
The future of FDE work seems closely related with all work around evals/posttraining/RL envs.
FDEs are effectively responsible for the following:
1. Define the business problem.
2. Codify the business problem into an eval rubric and environment.
3. Hillclimb the environment and output an agent/agentic workflow that solves the business problem.
Right now the process of (3) is quite manual - historically FDEs spend hundreds of hours creating bespoke software/workflows that solve the problem.
But assuming intelligence is abundant, they can effectively offload (3) to some automated optimization process. This includes RL on the model layer, and using Claude Code/Codex to optimize the harness/workflow.
Then the FDE responsibility shifts from implementing the task to defining the right goals and outcomes. In other words, they have access to /goal, and their job is more around making sure the goal, environment, and evals are correct vs. the tactical implementation details.
Fran Algaba retweeted
Ryan Dahl just launched celld: self-hosted, distributed Durable Objects.
One SQLite database per cell. S3-compatible storage. No separate database cluster or control plane.
I looked at the architecture, tradeoffs, and where I’d use it:
flaviocopes.com/celld/
great read on mcp design. also good monitoring around it is needed to reach these conclusions and many solutions lack that.
engineer on our team wrote a short piece about the AX of mcps.
turns out, there are problems with just throwing a bunch of tools into a bucket.
raindrop.ai/blog/mcp-design
working at @HappyRobot means building agents that survive contact with real operations.
we study production traces, find where the agent or workflow broke, turn those cases into evals, improve the prompts, tools and control flow, then run the regression suite before shipping again.
that loop matters more than making a model look clever in a demo.
a production agent is the whole system around the model: context, tools, permissions, latency, fallbacks, observability and end-state verification.
we own all of it, then use what happens in production to make the next version better.
very excited to share the journey working along with pablo and the team on this.
We just raised a $150M Series C at a $1.2B valuation led by Prysm Capital and co-led by @eurazeo. 🦄
Two years ago we deployed one voice agent in a US freight broker. Today, we’ve deployed a full platform in over 150 enterprises globally across supply chain, energy and utilities, telecom, insurance and airlines. 5x growth since our Series B.
Some of those incredible customers include @DHLGlobal , @Kuehne_Nagel, @Naturgy, @Repsol, and @Uber.
We’re building Enterprise Superintelligence: the state where an organization's collective intelligence compounds as agents and humans learn from one another.
And it starts by deploying your first agent. Much like that first agent we deployed two years ago.
Thank you Prysm and Eurazeo. And to @a16z, @Base10Partners, and @ycombinator for doubling down.
Thank you to the 200 @HappyRobot folks who worked so hard to make this happen. If you need us, we’ll be heads down, building a world where things just work. ▶️
& Thank you to our strategic investors including including @DisruptiveKoch, @orange, T.Capital (@deutschetelekom), @Bankinter, @endeavor_global, @Kfundvc and Wave-X.
Fran Algaba retweeted
recommended reading. in a few systems i built over the past 2 years i ended up with this flow for a given task:
- start with full inference for all steps, observe behaviour, judge correctness
- find inference steps that can be replaced with deterministic steps
eventually, most inference goes away, leading to more stable yet capable systems that can do much more than deterministic systems alone.
Wrote about this intuition a while back. Really hard to scale as each problem has an specific approach that will make it reliable. Hopefully we can discover good patterns!
davidgasquez.com/reliable-un…
this is incredibly important for anyone building production agents, specially in the enterprise.
data drifts quite quickly, edge cases appear. agents need to be self evolving with a well designed feedback loop to create roi over time.
The creator of Pydantic explained how to make live agents smarter after they ship in 1 hour 21 minutes - better than $2500 agent evals bootcamps.
collect traces from your live agent -> score them with an eval -> run GEPA to auto-optimize the prompt -> deploy the winner -> repeat forever.
That loop is how Pydantic AI agents self-improve while running in production.
Pydantic AI + Logfire tracing + GEPA prompt optimizer + evals + feedback loops - that's the stack.
Watch and save it, then wire GEPA into your live agent this week.
great read. i don’t feel confident enough to give loops enough autonomy to ship production code yet. nevertheless, it’s a tooling problem and it will be solved in the coming months.
Some more thoughts on looping in coding agents. lucumr.pocoo.org/2026/6/23/t…
very recommended view. lots of great insights.
We're trying a new experiment at @cursor_ai - interviewing devs we admire.
I chatted with @oneill_c & @part_harry_ from @baseten about how they use coding agents. We discussed their current dev workflows & some predictions for the future.
Check it out below!
this is the direction many infra / platform companies are going for integrating ai in production. the reality is that it also creates more value than offering plain subscriptions. this trend will only accelerate over time.
Services are the future. Today we launched Ramp’s AI services motion.
It's easy to buy an AI subscription. It's hard to transform your company to actually run on agents.
Here’s our entire strategy.
1) Why now
Services are the new software (Sequoia)
Human labor TAM >> software license TAM. The market is bearish on seats and subscriptions.
Every enterprise AI company is doing this -- the labs have poured billions into services partnerships and their own deployment functions. Superintelligent models alone are not enough.
Palantir proved this is a strong business model: deeply embed engineers, build on top of a powerful platform, and customize extensively.
2) The real problem
Companies want AI. But the gap between "we have AI tools" and "agents run our workflows and we spend way less time" is enormous.
What we've found across over 50 companies we engaged with: agents start replacing real work when there is: complete data, read/write access across systems, agent-friendly policies. Most big companies struggle because:
- processes live in operators' heads
- dozens of disconnected systems (legacy ERPs, endless one-off excel sheets, etc.)
- archaic software with poor or no API access
Good data in the right place is a hard prereq to working agents.
Also, vibing in localhost ≠ a production system your enterprise can rely on. You still need hosting, ci/cd, observability, feedback loops, good interfaces. And taste to know what's even worth automating.
Everyone has a bulldozer, but most jobs just need a shovel pointed at the right spot.
What companies usually need is to be made agent-friendly. That's exactly what we do.
3) What we do
We focus on what Ramp does best -- finance.
And we embed FDEs that:
-> understand your problems
-> identify high-leverage, high-impact workflows that fit agents
-> scope the solution
-> connect your data
-> capture your context
-> deploy agents and often bespoke software for humans to collaborate with them
-> drive the business metrics that matter
Discovery and scoping are crucial. Building is easier than ever and thus judgement about what to build is more important than ever.
We're not a generic AI services arm, we're finance domain experts. Across the spectrum of financial operations, we help companies find and frame the problems worth automating -- similar to the taste a founder has in choosing which problems are worth solving (ex-founders make great FDEs).
Here’s the stack we deliver:
- Production infrastructure. Shipping an index.html from Claude isn't the same as creating a repo, hosting in a cloud service, ci/cd, testing, setting up evals, managing memories and skills, adding feedback loops, ensuring uptime, incident management, etc. Agents don't one-shot production systems yet. Production software is hard -- we build, host, and run it for you in a single-tenant, dedicated cloud environment. Most operators don’t have the time, knowledge, or experience to do this e2e. We help abstract the low-leverage plumbing so they can focus on the essential parts of their jobs.
- Data connectivity. Most enterprises have data lakes, but data is often incorrect, stale, or entirely missing. And write interfaces vary dramatically. Ideally we can use MCPs or CLIs, but usually it’s poorly documented APIs, SFTP, manual uploads, and email.
- A context layer. Things people have done for years aren't written down, so an agent can't do them until we capture that context -- ranging from simple policies to complex decisions. This usually involves creating policy documents, shared agent memories, and skills.
- Evals and feedback loops. How you know an agent is doing a good job, and how it improves over time.
4) Why Ramp AI Solutions
We focus on finance because it’s the vertical we know deeply, have structural advantages, and are most differentiated:
- Data. 70k+ customers use our core product, over $200B in annual payments, years of vendor data, millions of transactions and bills monthly.
- Money-movement primitives and partnerships. Global money movement rails, partnerships with banks, Visa, Stripe, etc. You don’t want to vibecode international wires for bill payments.
- An intelligence layer on top: fraud detection from hundreds of millions of expenses, PO-to-invoice matching, state-of-the-art OCR, and fine-tuned models for accounting coding, spend routing, policy review, etc.
Unlike the labs, we’re not incentivized to sell tokens.
Ramp is an AI fiduciary and an impartial broker to deliver AI that is:
- model-agnostic -- we benchmark all the leading models (labs, open source) and fit the right one to each task
- and token-efficient by design
Our main incentive is business outcomes -- which is Ramp’s mission, to save our customers time and money.
I’m extremely bullish about our motion, and the broad industry growth of AI-native services.
If you're a finance leader trying to be more agent-native,
If you’re interested in joining our FDE team,
I’d love to talk 🙂
@perplexity_ai recently released search as code, basically codemode for search. the idea is to let the model generate a retrieval program instead of calling a search engine directly.
that program can fan out queries, filter sources, dedupe results, extract evidence, and control what reaches the model context.
i created the open source version of search as code for anyone to build on top of these primitives in their preferred harness. it supports @p0 as the core engine as well other providers and @wevm_dev curl.md for all the site rendering in markdown.
github.com/franalgaba/search…
people going crazy over loops with agents. i’m more biased towards recursion than loops for generating something valuable that is not slop in the current state of tooling.
this is a very cool way to optimize search. basically codemode applied to search engines.
Introducing Search as Code, our new search architecture for AI agents.
It writes Python that calls our search stack directly, instead of looping through function calls one at a time.
Available in the Perplexity Agent API, and now default in Computer.
research.perplexity.ai/artic…
Fran Algaba retweeted
someone hit me up about the new "claude dynamic workflows" feature, claiming "see, multi-agent works"
But really, the launch of this feature proves the exact point that I made back in June of 2025, along with @walden_yan, @tobi, @karpathy, and many others:
Deterministic workflows orchestrating small agent loops beats non-deterministic multi-agent or "agent soup" systems every dang time
everything is context engineering
the same model can behave very differently depending on the harness it runs in. they use RL to improve base models on the harness they are running with quite nice gains. also @badlogicgames pi outperforming nicely there!
really impressed with the latest @cursor_ai model composer 2.5. i downgraded my claude subscription for it. really impressed with the cost/performance of the model.
now in the 100$ sub in openai, the 20$ in claude and 20$ in cursor. the balance between all of them is awesome.