ScrutifAI

Use cases

What would you hand over?

Work that takes specialist software, hours of computing and files too big to upload. Say what you need in plain words. An agent runs the real tool, checks the result and sends back the finished files.

77 example jobs12 fields92 tools

Why not just ask a chatbot?

A chat can tell you how. This does it.

  1. The files won't fit.

    300 GB of raw reads. A folder of CT scans. A laser survey of a whole building. Chat uploads stop well short of that.

  2. The work needs the real software.

    Docking, meshing, alignment and load flow run in specialist tools, often for hours. A convincing paragraph isn't a finished run.

  3. You need the file, not the advice.

    Gerbers a fab will accept. G-code a mill can cut. Datasets that pass a regulator's rules. Work other people can open and rely on.

Medicine and healthcare

Scans in, measurements out.

Radiology and imaging research

Thousands of scans, and the same careful measurement on every one.

Tumour volumes across a trial

Segment the liver lesions in these 140 CT scans and track each lesion's volume over time.
You send
140 anonymised CT series (DICOM), about 60 GB
You get
Volume per lesion and timepoint, plus masks you can reopen in 3D Slicer
We check
Lesions matched across timepoints, and a sample of masks compared with a second method
  • 3D Slicer
  • nnU-Net

Also for pharma, contract research and veterinary oncology

Muscle and fat from routine CT

Measure muscle and fat at the L3 vertebra on these 2,000 abdominal CTs.
You send
2,000 anonymised CT scans (DICOM), about 400 GB
You get
Muscle and fat area per patient in a sheet, with an overlay image for each
We check
L3 slice confirmed on every scan, and outliers flagged for your review
  • TotalSegmentator
  • 3D Slicer

Also for oncology, nutrition research and sports medicine

A printable model from a scan

Turn this CT into a 3D-printable model of the pelvis for surgical planning and teaching.
You send
One anonymised CT series (DICOM)
You get
Watertight STL, ready for your printer
We check
Landmark distances on the model measured against the scan
  • 3D Slicer
  • Blender

Also for medical devices, veterinary surgery and medical education

Pathology

Slides too large to open on a normal laptop, let alone count by hand.

Cell counts across whole slides

Score Ki-67 across these 300 whole-slide images and give me the positive percentage for each.
You send
300 anonymised scanned slides (.svs), about 1 GB each
You get
Score per slide, and a QuPath project you can open and correct
We check
Threshold set on your control slides, with tissue folds and pen marks excluded
  • QuPath

Also for biobanks, pharma and toxicology

Neurology and neuroscience

Pipelines that run for hours on each participant.

Brain volumes from MRI

Get hippocampal volume and cortical thickness for all 250 participants in this MRI study.
You send
250 anonymised T1 MRI scans (NIfTI), about 10 GB
You get
Volumes and thickness per participant, with a quality image for each
We check
Failed reconstructions flagged, and the same software version used throughout
  • FreeSurfer

Also for dementia trials, psychiatry and concussion research

EEG cleaned and analysed

Clean these 80 EEG recordings, remove the eye blinks and give me alpha power by region.
You send
80 anonymised recordings (.edf), about 40 GB
You get
Cleaned recordings, power by region and group plots
We check
Every removed component listed, with the same settings for each participant
  • MNE-Python

Also for sleep clinics, neurotech and pharma

Rehabilitation, dentistry and devices

Scans and motion capture that end as a number or a physical object.

Walking analysis from motion capture

Run joint angles and joint loads for these 200 walking trials.
You send
200 motion-capture trials (.trc) and force-plate data (.mot)
You get
Joint angles and moments per trial, with the scaled OpenSim model
We check
Marker error reported per trial, and failed trials listed
  • OpenSim

Also for prosthetics, sports science, footwear design and workplace ergonomics

Study models from dental scans

Turn these 40 intraoral scans into labelled, hollow models ready to print.
You send
40 anonymised intraoral scans (.stl)
You get
Print-ready models and a print file for your machine
We check
Every mesh watertight, and each model compared with its scan
  • Blender
  • PrusaSlicer

Also for orthodontics, dental labs and veterinary dentistry

Virtual bend test for a bone plate

Run this bone plate design through a four-point bend in simulation. Where does it yield?
You send
STEP model, titanium grade and load case
You get
Stress results, yield load and solver files, for your design review
We check
Mesh refined until the peak stress stops changing
  • FreeCAD
  • CalculiX

Also for orthopaedics, sports equipment and aerospace brackets

Life sciences

From raw reads to results you can defend.

Genomics

Hundreds of gigabytes, long pipelines, and one wrong setting can spoil the lot.

RNA-seq from raw reads

Take these 48 samples from raw reads to a list of differentially expressed genes.
You send
48 paired FASTQ samples, about 300 GB, and a sample sheet
You get
Count matrix, gene tables, plots and the full pipeline report
We check
Read quality per sample, outliers flagged and every software version recorded
  • Nextflow
  • STAR
  • DESeq2

Also for biotech, crop science and veterinary research

Variants across a cohort

Call and annotate variants on these 30 whole exomes against GRCh38.
You send
30 de-identified exomes (FASTQ), about 200 GB
You get
Annotated VCF and a gene-by-gene summary
We check
Coverage and contamination per sample, with the pipeline tested on a reference genome first
  • GATK
  • Ensembl VEP

Also for rare disease, cancer research and livestock breeding

Is this one outbreak?

Assemble these 120 bacterial isolates and build the tree. Is this one outbreak?
You send
120 isolates (FASTQ), about 100 GB
You get
Assemblies, SNP distances and a phylogenetic tree
We check
Contamination checked per isolate, and the SNP threshold stated
  • SPAdes
  • Snippy
  • IQ-TREE

Also for hospitals, food safety, water companies and farming

Single-cell clustering

Cluster these 12 single-cell samples and label the cell types.
You send
12 samples of Cell Ranger output
You get
An .h5ad file to reopen, cell type maps and marker gene tables
We check
Doublets removed, batch effects compared, and each label tied to its marker genes
  • Scanpy

Also for immunology, cell therapy and cancer research

Drug discovery

The computing between a target and a shortlist.

Screen a compound library

Dock these 50,000 compounds against my protein and rank the top 100.
You send
Target structure (.pdb) and compound library (.sdf)
You get
Ranked hits and pose files you can open in PyMOL
We check
A known ligand redocked first to prove the setup, and problem compounds filtered out
  • AutoDock Vina
  • RDKit
  • PyMOL

Also for agrochemicals, veterinary medicine and academic labs

Does the binder stay bound?

Run 100 ns of molecular dynamics on my top five hits. Which stay bound?
You send
Five protein-ligand complexes (.pdb)
You get
Trajectories, stability plots and binding energy estimates
We check
Equilibration confirmed before measuring, with three independent repeats of each
  • GROMACS

Also for biologics, agrochemicals and enzyme engineering

Antibody structure predictions

Predict structures for these 30 antibody variants and show me where each one differs.
You send
30 sequences (FASTA)
You get
Structure files and an aligned comparison for each variant
We check
Confidence reported per residue, so weak regions aren't over-read
  • ColabFold
  • PyMOL

Also for enzyme design, vaccines and diagnostics

CRISPR guides across the genome

Design guides for these 200 genes and check each one for off-targets across the human genome.
You send
Gene list and genome build
You get
Ranked guides with off-target counts, in a sheet ready to order
We check
Off-targets searched genome-wide, not estimated
  • CRISPOR

Also for crop breeding, cell therapy and functional genomics

Lab imaging and cytometry

Instruments produce more files than anyone can look at.

High-content screen

Measure every cell in these 20,000 plate images and give me dose-response curves by compound.
You send
384-well plate images (.tif), about 150 GB
You get
Per-well measurements, curves, and overlays to spot-check
We check
Control wells scored for plate quality (Z-prime), and segmentation overlays sampled
  • CellProfiler
  • Cellpose

Also for toxicology, cosmetics testing and academic labs

Flow cytometry gating

Gate these 200 FCS files using my panel and report the population frequencies.
You send
200 FCS files, with your panel and gating strategy
You get
Frequencies per sample, with gating plots for every file
We check
Compensation checked against single stains, with the same gates on every sample
  • flowCore

Also for immunology, haematology and vaccine trials

Clinical research

Work a regulator's software reads before a person does.

Submission-ready trial datasets

Map these raw trial exports to SDTM and ADaM, and validate them for submission.
You send
Pseudonymised EDC exports, protocol and statistical analysis plan
You get
SDTM and ADaM datasets (.xpt), define.xml and tables for your statistician
We check
Run through the CDISC conformance rules, with every finding listed
  • R
  • CDISC CORE

Also for medical devices, vaccines and contract research

Dose modelling

Fit a population PK model to these 400 patients and simulate three dosing schedules.
You send
Pseudonymised concentration-time data from 400 patients (.csv)
You get
Fitted model, diagnostics and simulated exposure for each schedule
We check
Visual predictive check and bootstrap run on the final model
  • nlmixr2

Also for paediatrics, veterinary medicine and toxicology

Safety signals in FDA reports

Run a disproportionality analysis for this drug class across 20 years of FDA reports.
You send
The drug class and the reactions you care about
You get
Signal tables by drug and reaction, with the code to rerun them
We check
Duplicate cases removed, and results labelled as signals, not proven causes
  • openFDA
  • R

Also for medical devices, insurers and public health

Chemistry and materials

Calculated, not guessed.

Computational chemistry

Calculations that run for days and fail quietly if they're set up wrong.

Compare three catalysts

Calculate the energy barrier for this reaction on these three catalysts.
You send
Starting geometries (.xyz) and the reaction
You get
Optimised structures, transition states and an energy profile
We check
Each transition state confirmed by a single imaginary frequency
  • Psi4

Also for pharma process chemistry, fuels and speciality chemicals

Glass transition of a polymer

Simulate this polymer and estimate its glass transition temperature.
You send
Repeat unit (.mol) and chain length
You get
Simulation files and a density-temperature curve with the fitted value
We check
Repeated at three cooling rates and compared with a polymer of known value
  • LAMMPS

Also for packaging, adhesives and medical plastics

Materials and analytical labs

Screening and instrument data at a scale nobody does by hand.

Screen battery materials

Calculate voltage and stability for these 50 cathode compositions.
You send
50 crystal structures (.cif)
You get
Ranked table, with structure and output files for each
We check
Method benchmarked against published values for known materials
  • Quantum ESPRESSO
  • pymatgen

Also for electric vehicles, grid storage and catalysts

Untargeted LC-MS

Process these 400 LC-MS runs, align the peaks and identify what you can.
You send
400 runs (.mzML), about 200 GB
You get
Aligned feature table, with library matches and confidence levels
We check
Blanks subtracted, and drift tracked through the QC injections
  • MZmine

Also for food testing, environmental monitoring, forensics and pharma

Engineering and manufacturing

Files a factory can actually use.

Mechanical design

You need the part and the proof, not a lesson in the software.

Will this part hold?

Model this bracket from my drawing and check it holds 200 kg.
You send
PDF drawing, material and load
You get
STEP model and a stress report
We check
Mesh refined until the peak stress stops changing
  • FreeCAD
  • CalculiX

Also for medical devices, robotics and agricultural machinery

Heat sink airflow

Simulate airflow through this heat sink at three fan speeds. What's the peak temperature?
You send
STEP model, power and fan curve
You get
Temperature and pressure results, and the full OpenFOAM case
We check
Residuals converged, mesh independence tested and the energy balance closed
  • OpenFOAM
  • FreeCAD

Also for data centres, electric vehicles and medical electronics

3D scan to CAD

This part is discontinued. Turn my 3D scan into an editable CAD model.
You send
3D scan of the part (.stl)
You get
STEP model, and a map showing where it differs from the scan
We check
Deviation from the scan measured across the whole surface
  • CloudCompare
  • FreeCAD

Also for classic cars, museums, rail spares and prosthetics

CNC toolpaths

Program this part for my 3-axis mill, in 6082 aluminium.
You send
STEP model, machine and tool list
You get
G-code and a setup sheet for the operator
We check
Toolpath simulated against the stock for collisions
  • FreeCAD

Also for aerospace, medical implants and tool rooms

Aerospace and drones

Compare hundreds of options before anyone builds one.

Aerofoil sweep

Sweep these 40 aerofoils at my Reynolds number and rank them by lift to drag.
You send
Aerofoil coordinates (.dat), Reynolds number and angles
You get
Polars for each section and a ranked table
We check
Unconverged points listed, and the method checked on a section with known wind tunnel data
  • XFOIL

Also for wind turbines, racing yachts and motorsport

Concept aircraft model

Build a parametric model of this drone and estimate lift and drag across the flight envelope.
You send
Sizing sheet, and wing and tail dimensions
You get
OpenVSP model you can edit, and aero tables
We check
Panel results compared with a hand estimate at cruise
  • OpenVSP

Also for eVTOL, agricultural drones and university teams

Electronics and chips

The output has to pass someone else's rules, not just look right.

A board your fab will accept

Lay out this schematic as a four-layer board and give me the files my fab needs.
You send
KiCad schematic, board outline and your fab's rules
You get
Gerbers, drill files, bill of materials and placement files
We check
Design rule check clean, and Gerbers reopened in a separate viewer
  • KiCad

Also for medical devices, wearables and industrial sensors

Power supply worst case

Simulate this power supply across temperature and component tolerances.
You send
Schematic or SPICE netlist, with tolerances
You get
Worst-case ripple and efficiency, from 1,000 Monte Carlo runs
We check
Nominal case checked against your datasheet values first
  • ngspice

Also for automotive, medical electronics and renewables

Verilog to chip layout

Take this Verilog through to a GDS layout on the SKY130 process.
You send
Verilog source and timing constraints
You get
GDS layout, timing report and power estimate
We check
Design rule and layout-versus-schematic checks clean
  • OpenLane
  • KLayout

Also for university tape-outs, sensor start-ups and research labs

Energy and utilities

Will it work, and will it pay?

Renewables and storage

Yield and lifetime numbers that a lender will question.

Solar across a whole estate

Model solar yield for all 400 roofs in our estate, including shading from nearby buildings and trees.
You send
400 addresses (.csv) and your panel datasheet
You get
Yield and payback per roof, in a sheet and on a map
We check
Sample roofs compared with PVGIS, and unmatched buildings listed
  • pvlib
  • QGIS
  • PVGIS

Also for housing associations, schools, supermarkets and councils

Wind farm layouts

Compare these five layouts for wake losses and annual energy.
You send
Turbine positions (.csv), power curve and wind data
You get
Annual energy and wake loss by layout and by turbine
We check
Wind data tested for gaps, and the model checked on a published reference farm
  • FLORIS

Also for offshore wind, infrastructure investors and landowners

Battery ageing

Predict how this cell ages over 2,000 cycles on our duty cycle.
You send
Cell datasheet, cycling test data (.csv) and duty cycle
You get
Capacity fade by cycle, and a fitted model you can reuse
We check
Model fitted on part of your test data and tested on the rest
  • PyBaMM

Also for electric vehicles, grid storage and medical wearables

Grids and buildings

The studies that decide whether something can connect, or will pay back.

Can this feeder take the load?

Will this feeder take 2 MW of EV charging? Run the load flow for every hour of the year.
You send
Network model, and load and charging profiles
You get
Voltage and thermal results by hour, with upgrade options compared
We check
Base case checked against measured feeder data
  • pandapower

Also for data centres, bus depots and factories

Retrofit energy model

Model this office's annual energy use before and after a heat pump retrofit.
You send
Plans, plant schedule and 12 months of bills
You get
EnergyPlus model you can rerun, with savings by measure
We check
Baseline calibrated against your 12 months of bills
  • EnergyPlus

Also for hospitals, schools and landlords

Subsurface

Surveys and drill logs that only specialist software can open.

Seismic interpretation

Map the key horizons through this 3D seismic volume and tie them to the wells.
You send
3D seismic volume (SEG-Y), about 40 GB, and well logs (.las)
You get
Horizon surfaces, depth maps and an OpendTect project
We check
Horizons tied to the well logs at every well
  • OpendTect

Also for geothermal, carbon storage and offshore wind ground surveys

Geological model from drill holes

Build a 3D geological model from these 400 drill holes.
You send
Collar, survey and assay tables (.csv)
You get
3D model, cross-sections and tonnage by domain
We check
Some holes held back to test how well the model predicts them
  • GemPy

Also for mining, groundwater, tunnelling and geothermal

Buildings and construction

Drawings, models and the sums behind them.

Architects and contractors

Scans and models too heavy to read by eye, and drawings other people build from.

Laser scan to floor plans

Turn this laser scan of the building into measured floor plans.
You send
Point cloud (.e57), about 30 GB
You get
Floor plans as DXF and PDF, to scale
We check
Plan dimensions measured back against the scan at set points
  • CloudCompare
  • LibreCAD

Also for heritage surveys, factories and insurers

Clash check across models

Find every clash between the structure and services in these models.
You send
Structure and services models (IFC), about 2 GB
You get
Issues file (BCF) that opens in your BIM tool
We check
Clashes grouped, with location and tolerance stated
  • Bonsai

Also for hospitals, rail stations and data centres

Quantities from a BIM model

Take off concrete, steel and finishes from this model, level by level.
You send
Building model (IFC)
You get
Quantities sheet, with each line linked to its model element
We check
Totals reconciled to the model, and elements missing data listed
  • Bonsai
  • LibreOffice Calc

Also for cost consultants, developers and embodied carbon studies

Planning drawings for an extension

Turn my sketch and measurements into existing and proposed plans and elevations for my extension.
You send
Sketch, measurements and site photos
You get
Plans and elevations at 1:100, as PDF and DXF
We check
Every wall and opening against your measurements
  • LibreCAD

Also for shopfitters, schools and care homes

Structures and building physics

Analysis for someone qualified to review and sign.

Frame under wind and earthquake

Model this steel frame and run it under wind and earthquake loads.
You send
Frame drawings, sections and load cases
You get
OpenSees model, member forces and drift, for your engineer to review
We check
Reactions sum to the applied loads, and key members checked by hand
  • OpenSees

Also for bridges, stadiums and seismic retrofits

Daylight across a development

Run a daylight study on every room in this block of flats.
You send
3D model (IFC or .obj), with glazing and finishes
You get
Room-by-room daylight results and grid plots
We check
Sky model and surface reflectances stated, with one room checked by hand
  • Radiance

Also for schools, hospitals and offices

Land, environment and farming

Maps with the model behind them.

Surveying, forestry and quarries

Huge captures from the air, and one number someone will pay against.

Drone photos to a terrain model

Turn these 1,800 drone photos into an orthomosaic and a terrain model.
You send
1,800 geotagged photos, about 20 GB, and ground control points
You get
Orthomosaic (GeoTIFF), terrain model and point cloud
We check
Accuracy measured at your ground control points
  • OpenDroneMap

Also for construction, archaeology, solar farms and insurers

Stockpile volumes

How much material is in each stockpile on this terrain model?
You send
Terrain model (GeoTIFF) and stockpile boundaries (.shp)
You get
Volume per stockpile, with a map
We check
Base level stated, and each volume recalculated at a second resolution
  • QGIS

Also for mines, ports, waste sites and biomass plants

Trees and carbon from LiDAR

Count the trees and estimate carbon across 2,000 hectares from this LiDAR.
You send
Airborne LiDAR (.laz), about 80 GB
You get
Tree map, canopy heights and carbon by compartment
We check
Counts compared with your field plots
  • PDAL
  • QGIS

Also for carbon credits, power line vegetation and councils

Flood, air and ecology

Models that planners and regulators expect to see run.

Flood model

Model a 1-in-100-year flood on this river, with and without the new embankment.
You send
Terrain model, river flows and the embankment design
You get
HEC-RAS model, and flood depth maps as GIS layers
We check
Model tested against a recorded flood before the design is run
  • HEC-RAS
  • QGIS

Also for insurers, councils and rail and road operators

Air quality modelling

Model how this chimney's emissions spread over the town, using five years of weather.
You send
Stack details, emission rates, terrain and weather data
You get
Concentration maps, and results at each sensitive site
We check
Results compared with the legal limits, with every input listed
  • AERMOD

Also for factories, data centre generators and waste plants

Species from water samples

Identify the species in these 96 eDNA water samples.
You send
96 samples (FASTQ), about 20 GB
You get
Species table per sample, with confidence
We check
Negative controls subtracted, and matches checked against a reference database
  • QIIME 2

Also for water companies, aquaculture and gut microbiome research

Farming

Field-by-field decisions from imagery nobody has time to go through.

Crop stress through a season

Map crop stress field by field this season and give me a variable-rate file for the tractor.
You send
Field boundaries (.shp) and crop dates
You get
Weekly satellite maps and a prescription file (.shp)
We check
Cloudy images excluded, and each zone compared with your yield map
  • QGIS
  • Sentinel-2

Also for crop insurers, commodity traders and agricultural lenders

Plant counts from the air

Count plants per hectare and map the gaps across these 40 fields.
You send
Drone flights over 40 fields, about 100 GB
You get
Plant counts and gap maps per field, as GIS layers
We check
Counts compared with your hand-counted quadrats
  • OpenDroneMap
  • QGIS

Also for vineyards, forestry nurseries and crop insurers

Transport, logistics and telecoms

Test the change before you make it.

Fleets and warehouses

Decisions with thousands of moving parts.

Routes for a fleet

Plan daily routes for 40 vans and 3,000 drops, with time windows.
You send
Orders, depots and vans (.csv), and six months of past routes
You get
A route per driver, a file for your telematics, and the saving against your current plan
We check
Every time window and van limit verified by a separate check
  • OR-Tools
  • OpenStreetMap

Also for home care rotas, waste collection and field service

Warehouse at peak

Simulate this warehouse layout at peak season. Where does it jam?
You send
Layout plan and six months of pick data
You get
A simulation you can change, with queues and throughput by zone
We check
Model reproduces last December's throughput before any change is tested
  • SimPy

Also for hospital A&E, airports and factories

Roads and networks

See what a change does before anyone digs, builds or climbs a mast.

Junction redesign

Simulate this junction redesign at rush hour. Do queues get better or worse?
You send
Road layout and traffic counts
You get
SUMO model, with queue lengths and delays before and after
We check
Baseline matched to your traffic counts first
  • SUMO

Also for councils, retail parks, stadiums and ports

Mast coverage

Map radio coverage for these 60 masts across the county's terrain.
You send
Mast sites, antenna data and frequency
You get
Coverage maps as GIS layers, with the gaps marked
We check
Predictions compared with your drive-test readings
  • SPLAT!
  • QGIS

Also for emergency services, farm sensors and utilities

Finance and insurance

Whole portfolios, not a sample.

Insurance and lending

Every property on the book, run through the model.

Earthquake risk on a portfolio

Run our 50,000 insured properties through an earthquake hazard model.
You send
Property schedule (.csv), with construction types
You get
Expected losses by region and return period, with an OpenQuake job to rerun
We check
Geocoding rate reported, and hazard compared with a published map
  • OpenQuake

Also for reinsurers, banks and infrastructure owners

Flood risk on a loan book

Score all 20,000 mortgages on this book against the flood maps.
You send
Property addresses (.csv)
You get
Flood band per property, and totals by postcode
We check
Every address matched or listed, and a sample checked on the map
  • QGIS

Also for insurers, housing associations and councils

Investment and treasury

More data and more scenarios than a spreadsheet will take.

Backtest on tick data

Backtest this strategy on five years of tick data, with fees and slippage.
You send
Strategy rules, and your tick data or a market data source
You get
Full trade log, performance report and the backtest project
We check
Look-ahead bias tested, and results confirmed on held-back years
  • LEAN
  • Market data

Also for commodities, crypto desks and family offices

Stress test an options book

Reprice this options book under 500 market scenarios.
You send
Positions and market data
You get
Profit and loss per scenario, with the worst cases broken down
We check
Book reprices to market before any shock is applied
  • QuantLib

Also for corporate treasury, energy trading and pension funds

Ten years of filings

Pull revenue and segment figures from 10 years of filings for these 3,000 companies.
You send
Company list
You get
A database you can query, with a link to each source filing
We check
Totals tied to the filing, and restatements flagged
  • SEC EDGAR

Also for private equity, consulting and academic research

Science and research

Results someone can reproduce.

Physical science

Raw observations and model output too large to handle casually.

Telescope frames to a light curve

Calibrate these 2,000 telescope frames and give me a light curve for this star.
You send
Raw and calibration frames (FITS), about 60 GB
You get
Calibrated frames and a light curve, with the pipeline to rerun
We check
Comparison stars used to confirm the signal isn't instrumental
  • Astropy

Also for exoplanet surveys, satellite tracking and university teaching

Climate projections for a place

Extract and bias-correct climate model projections for this region to 2080.
You send
Region boundary and the variables you need
You get
Corrected data (NetCDF), with charts by scenario
We check
Bias correction tested against observed weather, and the spread across models shown
  • CDO
  • xarray

Also for insurers, water companies, farming and property investors

Research methods

Checking the work means running the work.

Does this paper replicate?

Rerun this paper's code and data. Do the tables come out the same?
You send
The paper and its replication package
You get
Published and reproduced results side by side, with a working environment
We check
Every mismatch traced to the code, the data or the software version
  • R
  • Python
  • Docker

Also for journals, pharma, policy teams and auditors

Any industry

The same job, in every industry.

Archives and records

Decades of recordings, paper and drawings that nobody can search.

Hours of audio to transcripts

Transcribe these 3,000 hours of recordings, with speakers labelled and every word time-coded.
You send
3,000 hours of audio, about 200 GB
You get
Time-coded transcripts you can search, with speakers labelled
We check
Error rate measured on a hand-checked sample, and low-confidence passages flagged
  • Speech to text

Also for call centres, courts, broadcasters and oral history

Millions of scanned pages

Make this archive of two million scanned pages searchable.
You send
Scanned archive (PDF and TIFF), about 2 TB
You get
Searchable PDFs and a database you can query
We check
Page counts reconciled, with accuracy measured on a sample
  • Document parsing

Also for councils, law firms, hospitals and libraries

Old drawings into a register

Convert these 18,000 old CAD drawings to PDF and build a register from the title blocks.
You send
18,000 drawings (.dwg and .dxf)
You get
PDFs, and a drawing register in a sheet
We check
File counts reconciled, and unreadable title blocks listed
  • LibreCAD
  • LibreOffice Calc

Also for utilities, rail, manufacturers and universities

Data at volume

Too many rows or pages for anyone to keep checking.

Competitor price tracker

Track 120,000 prices across these 40 competitor sites, every morning.
You send
Competitor sites and your product list
You get
Daily price file, matched to your products, with changes flagged
We check
Location confirmed on each page, not assumed from a success code, and blocked pages listed
  • Web scraping

Also for retail, travel, insurance and energy suppliers

Reconcile millions of transactions

Match two years of payments against the ledger and show me everything that doesn't tie.
You send
Payment and ledger exports (.csv), 14 million rows
You get
Reconciliation workbook and every unmatched item
We check
Totals tie to both source files, and every match rule is listed
  • DuckDB

Also for retailers, hospitals, payment firms and charities

Do you make one of these tools?

The same agents show us where software and APIs trip them up. If you sell one, we can show you where.

See the agent audit

Private waitlist

Which job would you hand over first?

We're opening to a small group first. Tell us the job or the software, even if it isn't listed here, and we'll start there.

These are examples of the jobs we're opening up, not customer results, and many aren't open yet. Health and life science examples are for research and for qualified people to review. They aren't a diagnosis, a treatment decision or a medical device, and patient data must be anonymised before it's sent. Anything that needs a professional's signature, such as structural sign-off, building regulations approval or legal advice, still needs a qualified person.