There is a particular kind of software project that everyone agrees must happen and nobody wants attached to their name. The code is old. The people fluent in it are retiring. The maintenance bill rises. But the program still moves money, replenishes stores, settles claims, or keeps a government service alive. Turn it off carelessly and the company discovers, in public, what the code was really doing.
The short version
- Imogen captures how a mainframe behaves, then rewrites and verifies it workload by workload.
- Customers include large banks, insurers, retailers, manufacturers, logistics firms, and public agencies.
- A public AWS listing shows a $100,000 monthly reference price; scoped contracts and cloud costs vary.
- The repeatable idea: make real inputs and outputs the acceptance test, then cut over in small pieces.
Mechanical Orchard was founded in San Francisco in 2022 to live inside that anxiety. Its premise sounds almost impolite in an industry fond of road maps: the documentation is not the specification. Neither is the COBOL. The specification is what the running system actually does.
That distinction is the whole company. Its platform, Imogen, watches data move through a legacy application, maps dependencies, helps generate clean replacement code, and feeds the same transactions through old and new versions. If the outputs diverge, the new code has not earned the right to ship. If they match, one workload can move while the rest of the mainframe carries on.
“The system in action is the specification.”Mechanical Orchard's operating principle
The $40 million lesson
The sharpest explanation of Mechanical Orchard begins with somebody else's failure. An unnamed Fortune 500 retailer had a seven-million-line mainframe estate supporting supply chain, merchandising, finance, stores, and inventory. It employed nine engineers to maintain it. Several were nearing retirement. The company hired a global systems integrator to translate COBOL into Java.
Three years and $40 million later, the converted program ran at one-tenth the performance of the mainframe. Nothing reached production. The retailer had purchased a large quantity of code and almost no new confidence.
Three years spent. One-tenth the required performance. Zero production releases.
Mechanical Orchard and the retailer chose the very batch process that had broken the earlier attempt. The team ingested source code and production data, interviewed users, mapped reads and writes, and turned observed behavior into tests. The replacement ran beside the original until full production runs reached behavioral equivalence. This was not a demo floating above the business. It was the process that had already embarrassed the previous project.
By September 2024, according to the company's case study, the retailer had retired one of its largest inventory and vendor-replenishment systems without reported disruption. Batch cycle time improved 265 percent and throughput rose 347 percent. The numbers matter, but the more transferable result was procedural: the team found out early whether a slice worked.
AI gets a chaperone
Plenty of tools can translate code. Hyperscalers can extract business rules. Consultancies can rehost an application or wrap it in APIs. Mechanical Orchard's claim to difference is the closed loop between generation and evidence. AI is useful for understanding tangled code and producing replacements quickly. It is not trusted to grade its own homework.
Imogen's Analyze layer maps code, data, and runtime relationships. Rewrite produces modern, idiomatic software rather than a line-for-line transplant. Orchestrate lets old and new workloads coexist while automated checks compare behavior. The 2026 release accepts analysis from both Google's Mainframe Assessment Tool and AWS Transform, then adds what those maps cannot fully contain: timing, dependencies, and odd production edge cases.
This is an old instinct in a fashionable coat. Mee founded Pivotal Labs, where teams practiced pair programming, test-driven development, and short releases. Mechanical Orchard co-founder Matthew Work came through leadership roles at Cognizant and Amazon. Chief scientist Kent Beck created Extreme Programming. The company may describe itself as AI-native, but its reflexes were formed before the current AI boom: write a test, make a small change, show working software, repeat.
During the pandemic, Mee's team replaced a state health department system. The client did not want to operate the new software, so the team kept responsibility for running and evolving it. That changed the proposition from “deliver a project” to “own the result.” Mechanical Orchard's mix of product licenses, engineering services, and operational support follows from that episode.
What it buys, and who buys it
The target customer is not a startup with a dusty admin page. It is an organization whose old software remains central because it works: a bank, insurer, retailer, manufacturer, automaker, logistics provider, healthcare operator, or public agency. Omni Logistics is a named early customer. Brazilian insurer SulAmérica has publicly discussed using Imogen on a critical system responsible for 30 percent of its mainframe consumption. Other customers are anonymized, which is unsurprising when the subject is the machinery inside a Fortune 100 company.
- Retail: inventory, replenishment, merchandising
- Insurance: claims, coverage, policy workflows
- Manufacturing: warranty and lifecycle systems
- Government: mission-critical public services
AWS Marketplace displays that figure for a one-month Imogen contract. Actual engagements are scoped by workload and can add cloud infrastructure. The listing also describes a free, NDA-covered proof of concept aimed at an initial replica and roadmap.
That is enterprise pricing for an enterprise problem. It also clarifies the business model: license the platform, implement it with Mechanical Orchard or a delivery partner, and run the new foundation in the customer's chosen cloud. Thoughtworks was the first global partner. Leidos extends the approach into U.S. federal agencies. Carahsoft provides a GSA Schedule route, while AWS and Google Cloud marketplaces shorten procurement.
A second case, and the useful copy
One retailer case can be a lucky story. A 2026 manufacturing engagement gives the method another test. A global industrial-equipment company needed to move an extended-warranty system to AWS. Working with Thoughtworks and Imogen, the team migrated four JCL batch jobs to Python on AWS Batch and moved three Db2 schemas to PostgreSQL. Hundreds of Java classes and more than a thousand SQL queries also went into production.
The lesson worth stealing is not “buy an LLM.” It is to move the acceptance test to the beginning. Pick one valuable but bounded workflow. Capture representative inputs and outputs. Build a parallel path. Compare both versions automatically. Keep the fallback. Put the narrow result into production before the program accumulates years of untested assumptions.
The approach has conditions. Teams need lawful access to representative production behavior. The old system must be stable enough to serve as a baseline. Leaders must tolerate a period in which two worlds run together, and they must prefer incremental cutovers to one theatrical launch. If the goal is to redesign broken business rules rather than preserve proven behavior, replication is only a starting point. It can tell you what exists; it cannot decide what the business ought to become.
Mechanical Orchard's bet is that most big companies have postponed this work because the usual methods demand belief long before they provide evidence. Imogen reverses the order. First the old machine testifies. Then the new one repeats the story. Only after their accounts match does anybody reach for the switch.