loading
Loading.Work
We don't run on testimonialswe don't have.
Real engagements, told straight: the context, the gap, what we did, and an honest outcome. Result numbers are gathered with each client before they go on the site. We publish numbers we can stand behind, or we don't publish them.
Each case is labeled: a client mission is paid work for a client; a Tsukumo build is something we built and run ourselves. The honest proof that we can run AI in production is that we run it on our own software first. Our products are on Products.
Client missions.
Paid work for clients: their product, their team, their repo.
Consulting · quant financeClient mission
Clareo Systems
Their own people running agents, safely, inside a correctness-critical environment.
- Context
- Clareo Systems is a quantitative finance firm. Correctness and risk control are non-negotiable, and the usual AI sales pitch, faster and more output, points the wrong way in a place where a confidently-wrong change is the expensive failure mode. The same properties that make agents valuable, speed and volume, are the ones a risk team is right to distrust.
- The gap
- A strong team using AI as a copilot, with the ceiling that implies: autocomplete in the editor, a human typing every consequential line. The jump to operating agents on real work is a skill and a process change, not a license, and in a controlled environment you can't bolt agents on top of the controls. They have to run inside them.
- What we did
- Agentic-development training for the team, then the agentic process and guardrails built for their environment, not a generic playbook: a human gate on anything irreversible, scoped permissions, and review at the right altitude. We upskilled the CTO to lead it, because the goal was their people operating agents on their standards, not a dependency on us.
Studio · real estateClient mission
A Swiss real-estate group
Marketing and product, both running on agents.
- Context
- A real-estate business that needed to scale content and marketing output (market notes, area guides, listings, editorial) and modernize its platform, without scaling headcount to match.
- The gap
- Manual content and marketing couldn't keep pace, and the platform wasn't built for an agent-first way of working.
- What we did
- Automated their acquisition marketing, stood up agentic blog-content creation in their voice, built an agentic-first platform, and ran GEO/SEO so the content gets found, including by AI engines.
Tsukumo builds.
What we build and run for ourselves. Our products are on the products page.
Self-proof · own ventureTsukumo build
Our own studio
We run what we sell, on our own operations and products.
- Context
- We are our own first client. We advise on running AI in production, and we run it on our own operations and products, which is the real test of everything we sell.
- The gap
- Real business operations (accounting, billing, payroll) and product development that a small team can't carry at the pace we hold ourselves to, without agents doing real work.
- What we did
- Built internal AI for accountants, an automated billing pipeline, and a payroll engine, and stood up autonomous dev teams that build and ship. The operating model we transition clients into is the one running our own business.
Outcome
AI-run accounting, billing, and payroll, plus autonomous dev teams shipping product.
Read the case →Studio · civic techTsukumo build
Open Truth
Applying agents to a problem most teams wouldn't attempt.
- Context
- Open Truth tackles verification of political claims, a domain where being right, sourced, and unbiased matters more than being fast or clever.
- The gap
- Claim verification at scale is beyond manual fact-checking, and a naive LLM that just answers is the wrong tool: it has to be sourced, checkable, and resistant to confidently-wrong output.
- What we did
- Built the algorithm plus the agent system that verifies political claims, designed so the output is grounded and auditable rather than an opinion. A novel AI system, not an off-the-shelf chatbot.