Life sciences R&D · United States, Europe and the United Kingdom · Operating since 2015
Anduril Limited is an independent AI software consulting company working only in life sciences R&D, with clients across the United States, Europe and the United Kingdom. We architect, build and harden AI systems for drug discovery and biomedical research — where a fluent answer and a correct answer are not the same thing, and where the difference has to trace back to a source someone can open.
Four things, offered separately or as one engagement. Most clients come to us after a promising prototype has stalled somewhere between a scientist's laptop and something the organization is willing to rely on.
Requirements elicitation with the scientists who will actually use the system, module boundaries, and deciding what it should do before any of it is written. That includes deciding whether a language model is the right instrument at all — for a meaningful share of R&D problems, it is not.
Agentic systems that reason across primary literature, target and compound databases, regulatory labels and trial registries, returning structured answers in which every claim resolves to a PubMed ID, DOI or database accession. Model-agnostic, and deployed so each program keeps its own data and its own trail.
Purpose-built tools for medicinal chemists, biologists, bioinformaticians and lab automation teams. The scientist describes the workflow; we build a working version quickly, argue about it with them, then convert it to production-grade code under test rather than shipping the prototype.
FAIR data practice, evidence standards, scope guardrails, and the organizational change that decides whether bench scientists use any of it. Technology is rarely the reason AI stalls inside a pharma or biotech research function.
We work in life sciences R&D only, and we work wherever the research does — biotech and pharma teams in the United States, across Europe, and in the United Kingdom. The engineering patterns below travel; two decades of domain knowledge — how experiments are designed, where data breaks down, what regulators care about, and why adoption stalls — is what makes them land.
A language model will produce a confident paragraph on a compound's binding mode, a target's off-target liability or a regulatory precedent whether or not the evidence supports it. We build the layer that decides whether to accept it: every citation resolved against its authority, every answer parsed against a schema, and weak results fed back as a tighter constraint on the next attempt rather than retried blindly.
Research teams rarely need a general assistant. They need a narrow tool that fits an existing workflow and produces an artifact a colleague, a committee or a reviewer will read closely. We work directly with the chemists, biologists and informaticians who will use the system, translate their process into a design, and build for the constraints that genuinely exist rather than the ones that look good in a pitch deck.
Some R&D questions are not questions, they are surveillance. We build systems that establish an empirical baseline from thousands of real trials or records, score each new entry against it, investigate only what deviates, and send nothing at all when nothing deviates. Deterministic code does the counting and scoring; the model is confined to interpretation and cannot contradict a computed fact.
Four stages. The middle two iterate. We do not treat the prototype as the product, and we do not start hardening something whose scientific purpose is still unsettled.
Framing is a separate, paid stage. If it concludes the system should not be built, or that the underlying data will not support it, that is a valid and useful outcome — and we will say so.
These are commitments rather than aspirations. They shape estimates, and occasionally they cost us work. In a field where a wrong claim can send a program down a dead end for a year, we would rather lose the engagement than soften them.
A system that fails quietly is worse than one that stops. Failure modes are detected, logged and surfaced at runtime. Where a check cannot be automated, it is documented as a known gap rather than left implied.
Data collection, counting and scoring belong in ordinary code that can be read and tested. The model reasons over what the code produced and is not permitted to contradict a computed fact.
When the evidence base is thin — a novel modality, a sparsely published target, an investigational compound with three PubMed records — the honest output is a recorded gap. Systems we build are designed to decline, and are tested on whether they decline where they should.
Source, tests and documentation transfer to your team. Deployments are isolated per program, so an unpublished assay set or a private target list stays inside your boundary. We are not trying to become a dependency.
We will not cut a corner in the build to hold a date. If a date is genuinely immovable, we reduce scope openly rather than ship something that fails its first review by a medicinal chemist.
Engagements run remotely by default, with working hours arranged around your team rather than ours — US East and West Coast, continental Europe, and the UK. We travel for the parts that genuinely need a room: framing workshops, lab visits, and hand-over.
Risk, missing information and the case against a proposal go into the same document as the recommendation. Clients get our honest read, including when it is inconvenient.
What we are not. We are not a CRO, not a regulatory affairs consultancy, and we do not validate systems for GxP or clinical use. Where an engagement runs into those, we build to hand off cleanly to the people who do, and we say so at the framing stage rather than at the end.
Anduril Limited is led by Dr Raminderpal Singh, an AI engineer and entrepreneur who has spent twenty years in life sciences — building depth in the industry's priorities, its risks, and its well-earned resistance to change — alongside earlier careers in semiconductor manufacturing and data infrastructure.
The foundation is systems engineering: understanding how complex parts fit together, where the friction sits, what the boundaries between components should be, and what it takes to make something work reliably at scale. That discipline has become more important with AI, not less. The speed at which an idea becomes running software has changed completely; momentum without structure produces fragile systems, hallucinated findings and untestable code.
He has run a genomics computing partnership at a major corporate research division, led market and product strategy for a life sciences data infrastructure company, and founded and led a business building causal-inference platforms for drug discovery. Earlier, he directed operations research inside a 300mm semiconductor fab. He holds twelve granted patents, has co-authored two technical books, and publishes regularly on AI in drug discovery.
Engagements are led personally, and scoped so that the client is talking to the person doing the work. Where a project needs additional hands, they are brought in under the same standards described above and named to the client in advance.
The most useful first conversation is usually a specific one: a screening or in-silico workflow that is not working, a prototype that has stalled on the way to production, or a scientific claim you need to be able to defend. Tell us what it is and we will tell you honestly whether we are the right people for it.
We work with clients across the United States, Europe and the United Kingdom. Written enquiries can also be sent to the registered office address below.