At an automotive technology company, the cloud had become a habit. Developers used remote virtual machines even for work their own computers could handle. Each machine was defensible in isolation. Together, they made an expensive default. Factored’s assignment was wonderfully unglamorous: work out which machines actually needed to exist.
- Embedded engineers build and maintain clients’ AI and data systems.
- Latin American talent and U.S. working-hour overlap underpin the model.
- The useful lesson: inspect the bottleneck before buying more technology.
The machines that could go home
Factored assessed utilization, standardized local development, introduced controls on provisioning, and moved suitable workloads in phases. It kept remote computing for jobs that needed it. The company reports 80% fewer development VMs and a projected 70% reduction in annual infrastructure costs. Those are different measures: a smaller machine fleet and an estimate of future spending.
The result turns the usual technology sales pitch inside out. An engineering services company can create value by removing infrastructure. The method is available to any engineering manager: inventory the work, separate genuine requirements from inherited customs, document a dependable alternative, and migrate gradually. Local machines still need enough capacity; computationally demanding work still needs somewhere else to run.
Illustrative index, normalized to 100 before the project
*Projected costs. Company-reported case results.
An export with no shipping container
Factored began in 2019 with Israel Niezen and co-founding advisor Andrew Ng. Its original problem was the shortage of AI and data engineers. Niezen saw an opportunity to connect Latin American technical talent with companies needing it. In a 2022 AI Fund interview, he described an acceptance rate of roughly 2-3% for incoming cohorts. That was a historical admissions claim, rather than a permanent guarantee about every engineer.
Selection is only the entrance. Factored describes live coding, systems thinking, code quality, communication assessment, and continued training. Its engineers join client environments and working hours. The commercial proposition is accumulated familiarity: someone who understands why a pipeline behaves strangely, which stakeholder owns a metric, and what a Friday deployment might disturb.
“For every cohort we bring on, the acceptance rate is around 2-3%”Israel Niezen · AI Fund interview, 2022
This places Factored between internal hiring, specialist consultancies, and engineering talent networks. A company can hire permanent employees, contract a project, or buy additional delivery capacity. Factored combines embedded specialists with organizational support. The distinction to examine when buying is who owns architecture decisions, delivery, and maintenance after launch.
The numbers have to survive the move
Its services cover AI applications, machine learning systems, statistical modeling, pipelines, and analytics. The customer might need a forecasting model, but also the data architecture that feeds it and the dashboard that makes its output useful. Factored’s partner roster includes Anthropic, Databricks, AWS, and Fivetran, putting its engineers among the tools enterprises already use.
Muck Rack provides a named example. In Factored’s account, the communications platform modernized legacy Snowflake tasks into dbt and Dagster workflows while moving permissions into Terraform managed through Git. The team refactored 43 tasks into 46 modular dbt models. Factored reports numerical parity before cutover and no production downtime. The engineering achievement is that the numbers survived the move.
Another enterprise needed help onboarding people to its data platform. Factored built an assistant that routes questions among environment setup, productivity, and engineering standards. It recorded 350 technical conversations in three months. That establishes use; proving time savings would require another measurement. The architecture’s practical idea is to send each question to the relevant knowledge and capability.
In healthcare, Factored describes joining fragmented clinical sources into consistent research datasets. In payments, it describes anomaly and time-series models that prioritize suspicious transactions. These customers share a basic difficulty: useful information is scattered, inconsistent, or buried in noise. Engineering makes it available in a form somebody can act on.
For a buyer, those examples suggest a useful starting brief. Pick a workflow with an owner and an observable outcome. Decide what must remain correct during a change, who can grant system access, and how the team will recognize progress. A migration can be checked against existing outputs. An assistant needs evaluation of its answers as well as evidence that people use it. These are practical questions to bring to an initial conversation.
The diagnosis was the bug
A more revealing example concerns an AI coding assistant. Factored describes a team investigating a dependency problem. The assistant blamed a supposed change in Pydantic 2.11 and proposed a custom validator. The explanation sounded plausible. The underlying claim was fabricated; further investigation found a dependency compatibility mismatch involving the Anthropic SDK.
What changed the investigation was checking the technical claim against evidence. A working patch would have been a poor substitute for a correct diagnosis. The transferable practice is simple: verify the claimed behavior change before modifying the application. An embedded engineer’s judgment becomes especially valuable when a confident tool sends the team toward unnecessary work.
A band called The Fellows
There is another side to the company’s engineering identity. Factored’s research portfolio includes contributions to The People’s Speech and DataPerf. Its culture pages describe continuing development and an annual gathering called Nerdapalooza, complete with an employee band, The Fellows. A distributed company still has to arrange occasions for people to know one another.
The commercial calculation belongs beside that human picture. Buyers should compare fees, onboarding time, and internal management effort with permanent hiring or another services firm. The benefit depends on the work: a continuing production backlog can reward retained context. A small, settled task may call for a smaller engagement. Client access, clear ownership, and available colleagues remain necessary ingredients.
Factored’s most instructive proposition is accountability at close range. Put engineers beside the people using the system. Let them learn the awkward details. Give them responsibility for what happens after deployment. Then judge the work by what the business can run, understand, and afford.