AI agentic workflows for scientific instruments

Software that operates your scientific instruments.

You describe the experiment in plain words. It reads the instrument's screen, sets the parameters, runs the procedure, watches the output, and tells you what it found.

The instrument can do it. The bottleneck is who operates it.

1.

Understanding what the instrument can do

Most users run the same three protocols because the manual is a thousand pages and the expert who set things up has graduated. The instrument's full capability goes unused.

2.

Running it, checking it, fixing what goes wrong

Setting parameters, babysitting the run, recognising bad output, adjusting mid-experiment — this takes a trained operator's full attention and most of their working day.

3.

Training the next person to do the same

Procedures live in heads and lab notebooks. When people leave, the knowledge walks out with them. Every new student starts from scratch.

Not a chatbot. An operator with deep context.

The agent sits in front of the instrument's screen the way an expert operator would. The difference is what it can hold in its head at once — and that it never forgets.

Every run it performs adds to its knowledge. Every manual it reads stays available. Every failure it logs becomes a diagnostic it can draw on next time.

Agent context Active
Reasoning
Plans multi-step procedures. Decomposes your request into instrument actions, checkpoints, and decision points.
Workflow library
Validated procedures the system can call by name. Each reviewed before it joins the library. Grows with every instrument it operates.
Run history & diagnostics
Every previous run, its parameters, its output, and what worked or failed. The institutional memory that usually leaves when people do.
Instrument knowledge
Manuals, specs, probe configurations, calibration records. Retrieved on demand, not memorised — always current, always complete.
Scientific literature
Published methods, application notes, vendor documentation. The agent reads the papers so the operator doesn't have to.

Five parts that read, plan, drive, learn, and stop.

The system attaches to the instrument's existing control software. Nothing is added to the hardware. The operator interface stays the same; another set of hands now uses it.

A trained operator turns a request into a plan, drives the console, reads the screen, recognises trouble, and decides when to stop and ask. The system does the same things, in roughly the same order.
A.

Reasoner

Reads your request, looks up the sample and instrument, and writes a plan: which procedure, what parameters, in what order, with what checks.

B.

Instrument interface

Drives the instrument's native software the way an operator would: clicks menus, types parameters, reads the same status messages. No vendor integration required.

C.

Knowledge base

Structured record of instruments, samples, and procedures the system has seen — including which combinations have worked and which have failed.

D.

Workflow library

Validated procedures the system can call by name. Each procedure is reviewed before it joins the library. New ones are added as the system encounters new instruments.

E.

Safety layer

Watches every action against a permissions tier. Anything that could affect the hardware stops and asks a human first. Defaults are conservative.

Want to see it run?

Write to us and we'll show you the system operating an instrument — your experiment, or one of ours.

Request a demo

What it's allowed to do, and what it has to ask first.

Permissions are explicit. Each action the system takes is checked against a tier, set by the facility manager and visible in every run log. Nothing happens silently.

Defaults are conservative: a new instrument starts at Tier 1. Higher tiers are unlocked only after a facility manager signs off.

Tier 0Observe Reads the console and the output. Takes no action on the instrument. No auth needed
Tier 1Standard runs Executes procedures from the library on approved samples. Reversible parameters only. Pre-approved
Tier 2Adjust parameters Modifies acquisition parameters within bounds the facility has set. Stays inside the envelope. Pre-approved + log
Tier 3Hardware-affecting Probe change, calibration, anything that could harm the instrument or invalidate downstream work. Human required
NMR spectroscopy

Our first instrument: the NMR spectrometer.

Nuclear magnetic resonance is one of the most powerful and most operationally demanding instruments in science. A 600 MHz magnet that runs nine to five is running at a third of its capacity.

We built the agent here first — shimming, locking, running pulse sequences, processing spectra, diagnosing problems — because if it works on NMR, it works on anything with a screen and a procedure.

RUN №1247 Acquiring
sampleubiquitin · 0.8 mM · D₂O
sequence2D ¹H–¹⁵N HSQC
field600 MHz
requested"check folding, overnight"
tier1 · standard, reversible
progressscan 164 / 256
started 22:14:08 ETA 06:31

The technical case, in the open.

Preprint

Agentic operation of NMR spectrometers

Pervushin K., Filippov M., et al. · bioRxiv, 2026
Conference talk

ENC 2026 — Experimental Nuclear Magnetic Resonance

Presented by Konstantin Pervushin
Conference talk

Asilomar Conference on NMR

Presented by Konstantin Pervushin
Tutorial

AI4X 2026 — Agentic instruments tutorial

Half-day with vendor and academic panellists
Let's meet there →

Built where the instruments live.

Co-founder · Instrument science

Konstantin Pervushin

Structural biologist. Thirty years operating spectrometers, from TROSY to in-cell NMR. Faculty at NTU, Singapore.

The reason the system gets the experiments right.

Co-founder · AI systems

Mikhail Filippov

PhD in physics. Built ML systems used by NASA and central banks; founded an AI engineering company before this. Faculty at NUS College, Singapore.

The reason the system gets the software right.

We're a small team in Singapore. If this is your kind of problem, we'd like to hear from you.

Get in touch →

Let's put it on one of your instruments.

We're taking on a small number of pilot facilities through 2026. A pilot starts with a single instrument, an agreed list of procedures, and Tier 1 access only.

You keep your data, your software, and your hardware untouched. We set up, train on your standard procedures, and step back.

Prefer to talk first? m@spinogryz.com · or book a 30-min call →
We reply within two working days.

Thanks — we'll be in touch.

Expect a reply within two working days.