For companies with a process that eats people's time. We build a brain that holds what your company knows and an operating system of agents that does the work. Live in production, measured in dollars and hours.
$14B+ in loans processed on a marketplace we built. ~30% more sign-ups reaching live after a sales-led flow went self-service.
in loans processed on RealAtom, a lending marketplace built on $5M of seed capital.
product and engineering org built from zero at a real estate platform valued above $1B, SOC 2 included.
lenders brought onto that marketplace, at 51 to 65% twelve-week retention.
accuracy on a machine learning prediction engine built for a venture-backed client.
Describe your software or a task your team repeats, in your own words. A model reads it and sets up the right calculator with your numbers. Or fill it in by hand below.
The brain is what your company knows. The operating system is what it does with it. Either one alone is a demo.
You already know how to win. That knowledge sits in decks, tickets, chat threads and three people's heads. We pull it into one structured memory that your people and your agents both work from.
A brain that only answers questions is a wiki. The OS does the work: your team runs skills by name, agents close tasks end to end, and loops run every morning and stop themselves when the output is wrong.
Multi-agent systems for the work that eats your team: intake, document review, compliance checks, content, reporting. A human reviews where the stakes justify it, and nowhere else.
Strategy, architecture and build for AI-native SaaS, from first sketch to paying customers. We have taken products from seed funding to billions in transaction volume.
The handoffs, re-keying and status chasing between your systems, removed. Event-driven, observable, and boring in the best way.
Replace a manual, people-heavy motion with a self-service one. We have done it for sales, onboarding and back-office review inside regulated, high-trust businesses.
Classifiers and fine-tuned models for the cases where a general model is too slow, too costly or too loose. One recent build: a 1.9B-parameter model fine-tuned on 150,000 examples across 176 categories, after a smaller one kept hallucinating.
Hands-on training for executives, operators and small teams. Nobody watches slides. Everyone builds something they will use on Monday.
We are not tied to a vendor. The right model is the cheapest one that passes your evals, and the right framework is the one your team can run after we leave.
Three builds running in production at a global two-sided marketplace. The client is confidential. The numbers are not.
Each step has a fixed price, agreed before it starts. There are no retainers and no hourly billing.
You bring one process that hurts. We ask what it costs today, in hours and in dollars.
30 minutes · freeWe sit with the people who do the work and map where the time goes. You get a ranked list of what to automate, with a number on each line.
About 2 weeks · fixed priceOne agent, end to end, on your real data and infrastructure. It ships with evals, so you can see how often it is right.
4 to 6 weeks · fixed priceWe widen what works, train your team to run it, and leave you the code, the prompts and the runbook.
Scoped from resultsHands-on sessions built around something your people ship by the end. Open courses run through Masha Sharma Academy, a Citrusworks client.
See the academyA free MCP server that measures how you write, with five practical uses: checking AI drafts, writing across channels, diagnosing your style, ghostwriting and house style.
October 12, 2026 · 2 min readTypeSafe's Jev and OpenAI's Decisions API both return a typed decision with a probability instead of text. What each one is, what is still unproven, and what to do about it.
September 23, 2026 · 2 min readFrontier, open or fine-tuned? Choose the cheapest, fastest model that passes your evals. Here is how to run that decision.
A corporate brain is one structured memory of what your company knows: its decisions and the reasons behind them, its playbooks, its metric definitions and what customers said. It is written so that both people and AI agents can read it.
An AI operating system is the layer that turns that knowledge into finished work. It is made of skills your team runs by name, agents that complete tasks end to end, scheduled loops with stop rules, and evals that score every output before it ships.
The first call is free. The diagnostic and the first build are each a fixed price, quoted before the step starts, so you know the number before you commit. If a step does not earn the next one, you stop.
About two weeks for the diagnostic, then four to six weeks to a first agent running in production on your data.
It stays in your systems. We work inside your environment and your accounts, under your NDA, and your data is not used to train anyone's models.
You do. The code, the prompts, the fine-tuned models, the eval sets and the runbook are yours, and your team is trained to run them.
One owner with the authority to change the process, access to the data it runs on, and time with the people who do the work today.
It takes the repetitive part of the job. People keep the cases where a mistake is expensive, and those are the cases they were hired for.
Then we say so on the first call or in the diagnostic. Some problems need a simpler process or a plain script, and that is cheaper for everyone.
Every agent ships with an eval suite. You see the score before anything reaches a customer, and you keep seeing it after we leave.
Tell us which workflow costs you the most hours or the most money. We will tell you whether the answer is a brain, an agent, or neither.
Citrusworks started in Arlington, Virginia in 2010 as a software studio. Today it does one thing: it puts AI to work inside companies. Senior operators only. The people who scope the work are the people who build it, and nothing is handed to a junior team.