Software Projects in Oxford Often Begin as Research Code
A large share of the custom development commissioned in this city starts from something that already works, sort of. A script that produced a paper. A model a postdoctoral researcher wrote and then left behind. A prototype built to win funding, now being used by real people it was never designed for. The buyer is rarely starting from a blank page; they are asking a development firm to turn a working experiment into a system other people can rely on.
That is a specific kind of project and it is priced badly by firms who have not done it. Research code optimises for getting an answer once. Production code optimises for getting the same answer repeatedly, for people who did not write it, on data that arrives messy, with a record of what happened. Moving between those two states is rarely a refactor. More often it is a rebuild that keeps the logic and discards the structure, and the honest suppliers will say so in the first meeting rather than quoting a cleanup. Ask every candidate directly whether they would keep the existing code, and ask them to justify the answer against maintenance rather than against sunk effort.
Background and Foreground Intellectual Property Need Separate Clauses
Where a university, a spinout, an academic collaborator or an institute is anywhere near the project, ownership is more layered than a standard development contract assumes. Existing algorithms, datasets, models and materials brought into the work are background intellectual property, and they usually belong to the institution or to a licensing entity rather than to the person handing them over. Anything created during the engagement is foreground, and it needs assignment terms that are compatible with whatever licence governs the background.
Write the two separately. Establish who owns the background, on what licence it is contributed, and what happens to it if the collaboration ends. Then make the foreground assignment explicit, binding on subcontractors and individual developers as well as the signing company, and surviving termination. Check whether a technology transfer office needs to approve the arrangement before work starts, because obtaining that approval afterwards is considerably harder. If the supplier brings its own framework or accelerator, secure a perpetual, transferable licence to keep using and modifying it, or exclude it from the build. A spinout that intends to raise investment should treat this as diligence preparation; a confused chain of title is one of the few technical findings that can genuinely delay a funding round.
Funder Conditions Reach Into the Repository
Grant funded work carries obligations that development firms meet for the first time when they meet you, and those obligations bite at the code level rather than at the reporting level. Many funders expect outputs to be made openly available, some expect software to be released under an open licence, and most require a data management plan describing where data lives, how it is described, how long it is kept and how it will be archived. Some require acknowledgement and persistent identifiers on released outputs.
Read the conditions before the technical requirement, because they constrain the design. An open release obligation is incompatible with a supplier that wants to retain proprietary components. An archiving requirement changes how you store data from the first week rather than at the end. Audit rights mean records have to survive the project. Put all of it in the brief, ask candidates whether they have delivered under funder conditions before, and make compliance an acceptance criterion so that it is somebody's job rather than everybody's assumption.
Validation and Reproducibility Change the Meaning of Done
In ordinary commercial software, done means the feature behaves as described. In research adjacent software, done usually means something stronger: that a result can be reproduced, that the version of the code and the version of the data which produced a given output can both be identified later, and that a reviewer or a regulator could follow the trail. That is a testing and engineering requirement, not a documentation one.
Ask candidates how they would version data alongside code, how they would record the parameters of a run, and how they would let someone rerun last year's analysis and get last year's answer. Ask what their automated test strategy looks like for numerical work, where a subtle change can produce plausible but wrong output that no interface test will catch. If the software will ever support a clinical, environmental or regulatory claim, raise validation expectations at design stage and agree who writes the evidence. Suppliers with scientific experience will welcome this conversation. Suppliers without it tend to reply that they follow best practice, which is not an answer.
Application Work and Platform Work Are Different Purchases
A lot of the demand in this city is described as app development, and that word covers two very different jobs. One is a contained application with a clear audience: a participant facing tool, a data capture instrument for fieldwork, an internal utility. It has a defined surface, it can be built quickly, and it is priced accordingly. The other is a platform that other systems depend on, with accounts, permissions, integrations and a lifespan measured in years.
Firms will quote for whichever you ask about, so be precise. If you need the first, do not buy the second, and consider whether application developers or web engineering firms are a better fit than a full software house, since that shape of work usually prices lower. If you need the second, judge the supplier on operations rather than on build: how they handle upgrades, monitoring, incident response, access control and the slow accumulation of data. Asking a supplier to tell you which of the two your idea really is, and why, is one of the more useful screening questions available.
Questions Worth Asking an Oxford Shortlist
Give every candidate the same document: the problem in plain language, the existing code and data and their provenance, the institutional and funder conditions attached, the systems that must connect, the commercial model you want, acceptance criteria including reproducibility expectations, and what support looks like after the grant or the phase ends. Ask who specifically will do the work, what they would do in the first month, and what they would refuse to promise.
Ownership of the practical assets matters here as much as the legal clauses: keep the cloud accounts, the source control organisation and the credentials in your own or your institution's name, with the supplier added as a collaborator. Verified profiles for software companies are listed on Edvido alongside communications teams working in the same city, and buyers who want to widen the field can review suppliers in the other research city or the capital's market. When the brief is ready you can put it to several firms at once and compare answers that were written against the same question.