A few spare minutes before dinner seemed harmless enough. Randall Julian opened an old project, fed some prototypes and data into an AI coding tool, and began experimenting. Seven hours later, he had built himself a compiler. Dinner had escaped. The original task, somehow, had acquired an entirely different job description.
He told the story himself in 2026. It is a pleasingly awkward admission from someone whose professional life involves helping computers make better use of scientists’ time. Even the person selling the machinery can get a sleeve caught in it.
Julian is the founder, CEO and chief scientific officer of Indigo BioAutomation, the Indiana company that develops software for laboratory data analysis. He is also a chemist, teacher and author. Across those jobs runs a persistent question: what must happen between a machine producing an answer and a person having good reason to believe it?
His career offers several versions of the problem. There is the computer too limited for an experiment, the instrument file another program cannot read, the spreadsheet quietly changing its contents, and the AI demonstration wearing clothes several sizes larger than its accomplishments.
A computer powerful enough to ask
During his doctoral work in the early 1990s, Julian noticed researchers using early personal computers to explore new instrumentation. Those machines imposed limits on the models they could attempt. He went looking for more capacity, using Cray supercomputers at Purdue and the Connection Machine at Los Alamos.
The point was to model instruments before building them. More computing power expanded the questions a researcher could ask. This was an attraction with consequences: Julian could see a substantial gap between the computers available and their everyday use in chemistry.
He earned his Purdue chemistry Ph.D. in 1993 and spent 14 years working in discovery chemistry at Eli Lilly. Informatics developed in his research group became the basis for Indigo. The computational side of the work eventually needed a company of its own.
There was a very ordinary nuisance underneath the technical ambition. Scientists spent time moving information among programs, cutting and pasting instead of investigating. Indigo developed automated analysis and visual interpretation to reduce that clerical burden. A laboratory could acquire sophisticated instruments and still spend an embarrassing amount of its day shuffling their paperwork.
Learning to occupy the corner office
Founding a software business did not immediately make Julian an experienced business leader. He recruited Raul Zavaleta to run Indigo, leaving himself room to concentrate on product development. The arrangement allowed the scientist to keep working on the part he knew while learning what the rest demanded.
In 2014, following an $8.5 million investment round, Julian succeeded Zavaleta as CEO. Zavaleta stayed involved as a consultant and director. Julian had accumulated years of experience helping operate the business by then. The change was a handover between colleagues rather than an origin story compressed into a promotion.
“It’s always been my intention to run my own company,” he said at the time.
Taking over also meant managing a change in how money arrived. Indigo had shifted from software licence fees to cloud subscriptions in 2012. Revenue came in more gradually under the new arrangement. The company had chosen that transition after a strong year, giving it resources to make the move.
For a scientist accustomed to asking whether a method works, the commercial question introduced another clock. A product could gain users while its revenue took time to catch up. The executive job required understanding both.
Getting the experiment out of the machine
One earlier project illustrates a different obstacle: scientific data can be perfectly useful and inconveniently trapped. In 2009, Indigo and Prosolia combined technologies in a software collaboration called FireFly. Its purpose included extracting data from proprietary systems and converting it for use in other software.
Both businesses had connections to Purdue. The collaboration brought together instrument expertise and data-conversion capabilities, allowing researchers to move information between tools and share it across organizations. The difficulty was practical enough to be understood without a chemistry degree: your collaborator needs to open the thing you have sent.
Julian also worked on the agreements that make such exchanges possible beyond an individual product. He previously chaired the Human Proteome Organization’s Proteomics Standards Initiative steering group and an ASTM committee on analytical data standards.
He is a coauthor of the 2012 mzIdentML paper, describing a common format for results from mass-spectrometry-based protein identification. The author list is a reminder that a standard is collective work. People have to agree on how to describe a result before their software can exchange it consistently.
There is little theatrical glamour in this. File formats seldom receive a standing ovation. Their usefulness becomes obvious when someone needs to reopen an experiment, compare results or hand the work to another group.
A lesson hiding in a spreadsheet
Julian’s teaching gives the same concerns a smaller stage. Alongside his work at Indigo, he maintains an adjunct chemistry appointment at Purdue and teaches short courses in data analysis. His public talks include a 2022 presentation on machine learning in mass spectrometry.
His 2025 book, R Programming for Mass Spectrometry: Effective and Reproducible Data Analysis, follows the work from handling data through statistical exploration, processing spectra and chromatographic signals, and applying machine learning. The title’s promise of reproducibility is central to the exercise.
The opening chapter contains an example with an almost comic lack of spectacle. A chemical registry number goes into Excel. The program interprets it as a date. A subsequent conversion produces a number with a different meaning. The spreadsheet has been helpful in precisely the wrong way.
Julian reproduced the example while writing. It is a teacher’s choice of detail: an error a reader can recognize, inspect and guard against. The machine requires no malice to damage the work. A convenient assumption will do.
The book uses R’s programming and reporting tools to keep analysis explicit and repeatable. In 2025, he was also scheduled to teach a two-day course in Protaras, Cyprus, with practical exercises covering data handling, noisy measurements and machine learning. The executive still had students to think about.
The demo and the work that follows
By 2026, AI-generated software offered a fresh version of the old question. Julian responded to coverage of an app built through a few prompts, distinguishing a prototype that operated on already processed data from a platform capable of the broader analytical work suggested by its description.
He acknowledged the accomplishment while questioning the label. That distinction matters in a field where data handling, identification, statistical analysis and validation involve different responsibilities. Showing one useful part does not establish that every other part has been built and tested.
“An architectural rendering can be beautiful, and it can be useful. But nobody can move into one.”
Randall Julian, 2026
His own research continued alongside the commentary. In June 2026, he coauthored a paper with Brian Rappold, Fan Yi and Stephen Master on Monte Carlo wavelet analysis for objective peak detection. The research addresses deciding whether apparent peaks in complex instrument data are statistically significant.
It is a concrete scientific counterpart to his broader concern about persuasive outputs. A visible feature can invite a conclusion. Establishing what the feature supports takes a method. Julian’s work keeps bringing him back to the space between those two moments.
The people around the calculation
Software has colleagues attached. When Indigo planned its 2017 move to Carmel, Julian described an office search shaped by employee input. Then-chief operating officer Sharon Mcilwaine gathered views about the space and location people wanted.
Access to running and biking trails mattered. So did nearby places for weekly team lunches. The chosen setting offered private areas and collaborative ones. A workplace for people solving difficult problems needed room for both conversation and concentration.
Those details are modest beside supercomputers and international standards, but they give the founder’s work a recognizable human scale. Someone must choose where the team sits. People will need lunch. A good calculation does not settle either matter.
Dinner, eventually
The AI detour brought Julian’s appetite for technical work into focus again. His response was to define completion before beginning, make room for delegation, and insist on independent checks that an agent cannot simply rewrite to suit its output.
The rules fit a career spent moving between scientific possibility and the conditions required to use it. He still wants computers to do more. He also wants the steps, the checks and the human decisions to keep pace.
The young researcher sought a machine powerful enough for the model. The experienced founder now writes down where the task should end. Between those habits lies much of his working life: curiosity with something useful to accomplish, and a growing appreciation for knowing when to close the laptop.