80.9%
Medical QA · historical aggregate
195/241 across batches · includes one retest
Nexus Platform — two product lines
Medical and agricultural AI,built on evidence.
Two dedicated product lines on one platform. Nexus Health OS helps doctors discuss clinical questions, interpret images and review reports in one conversation. Nexus Agriculture brings field photos, agronomic questions and farm records into a shared discussion.
The first product line
Nexus Health OS
From clinical questions to image review.
A familiar conversation, with the materials you need alongside it. Ask directly, add a file and continue the discussion without a separate clinical form.
Recorded research evaluations
80.0%
Literature QA · 40-question run
32/40 exact answers · 3 unscorable responses
82.0%
Medical exam questions
132/161 MedMCQA questions across five batches
83.3%
DDX+ reference-label coverage
5/6 fixed cases · mentions in free-text answers
0.879Dice
Brain MRI whole tumour
50-case sealed cohort · core 0.877 · enhancing 0.816
0.978Dice
CT lung-field research
Reference lung-field task · not nodule diagnosis
0.833AUROC
Pathology · grouped-tile study
11 evaluation patients · sampled tiles grouped by patient
154checks
Ultrasound engineering validation
154/154 checks of runtime and evidence handling
These are separate research tasks, cohorts and versions, not one production-system accuracy score. The first-run medical QA aggregate was 193/241 (80.1%); 80.9% includes a repeated batch. Reports retain methods, source records and qualification results. Research evaluation does not establish clinical approval.
- Clinical questions
- Work through symptoms, possible causes and next steps, adding clinical context as the discussion develops.
- Medical images
- Discuss findings in the images you upload, with the original view available throughout the conversation.
- Reports & laboratory results
- Read reports and laboratory results together, keeping the relevant details close to the clinical question.
- Treatment & medicines
- Compare treatment options, medication considerations, interactions and follow-up plans in context.
- Follow-up discussions
- Pick up a discussion where you left off. Earlier questions and attached files stay within reach.
- Private workspaces
- Keep your history in an individual account, with a choice of English or Chinese.
Nexus engineering
A shared foundation for medical intelligence.
Our platform connects medical file handling, structured evidence, specialist methods and clinical data standards. These reusable components let each discipline develop without losing its connection to the original examination.
- A common evidence structure
- Source references, observations, measurements and spatial evidence share a consistent representation. Findings can be traced back to the material that produced them.
- Coordinates that retain their meaning
- Image pixels, volume voxels, physical coordinates, slide levels and waveform timestamps remain distinguishable. A point on a display is not treated as a point in every examination.
- Specialist capability routing
- Components are registered by modality and intended use, with separate versions and qualification states. A new specialist method can be integrated without rebuilding the whole product.
- Structured clinical knowledge
- Knowledge-graph and medication components preserve source versions, clinical relationships and the context of supporting or conflicting evidence.
- Comparison across time
- Longitudinal components account for examination identity, acquisition differences and alignment before interpreting a change.
Clinical inputs
Images · signals · documentsUnified evidence
Sources · context · coordinatesSpecialist modules
Modality-specific methodsClinical interfaces
Conversation · clinical records
Specialist research
Specialist imaging, signals and genomics.
Nexus maintains a multidisciplinary research codebase, spanning radiology, pathology, optical imaging, physiological signals and molecular evidence. Each area has its own data handling, spatial requirements and evaluation methods.
These are research and engineering capabilities, not a list of functions enabled in the current conversational product. Availability and performance are assessed separately for each intended use.
View research figuresCT · MRI · PET/CT
Volumetric imaging
Three-dimensional source mapping, segmentation and physical measurements, with separate qualification for brain-tumour MRI and other specialist tasks.
- Source data
- CT, MRI and paired PET/CT examinations
- Reference system
- Source series, voxel indices and physical LPS coordinates
- Research focus
- Three-dimensional segmentation, measurements and examination alignment
Evaluation is part of the engineering.
Measure the actual task
Detection, segmentation, classification and measurement use different metrics. We do not reduce them to one system-wide “accuracy” figure.
Keep evaluation independent
Qualification checks use patient-separated public reference evaluations and declared thresholds, rather than treating a successful demo as a clinical benchmark.
Retain the evidence
Dataset versions, spatial results and failed baselines remain available for comparison. A weaker experiment does not become the default just because it runs.
Research in view
Spatial evidence, in view.
Nine figures show six brain-MRI views plus CT lung-field, retinal and pathology research outputs. They come from different research workflows. An example is not a cohort result or a claim that its specialist is enabled by default in the conversational product.
Model prediction beside the expert reference
Same patient, same slice: the sealed-evaluation prediction (left) against the manually delineated expert reference (right). Whole tumour Dice 0.958 · core 0.976 · enhancing 0.945, recomputed voxel-by-voxel on the spot.
Multi-sequence tumour atlas
T1, contrast-enhanced T1, T2 and FLAIR with a retained research prediction — whole tumour (cyan), tumour core (blue), enhancing tumour (red) from one inference.
Three-plane spatial mapping
Axial, coronal and sagittal views from one inference, delineating tumour subregions in true three-dimensional space.
Nested three-dimensional reconstruction
Whole-tumour point-cloud envelope, core shell and enhancing region rendered from the same research prediction to inspect three-dimensional geometry.
Volumetry panel
Measurements from this research prediction: whole tumour 170.1 mL / 106.4 mm, core 69.4 mL and enhancing region 47.6 mL. These describe one example, not cohort performance.
A retained prediction across sampled slices
Eight sampled slices around the same tumour from one retained prediction. Slice indices are shown; the sampling intervals are not uniform.
Chest CT lung-field segmentation
Research lung-field output over a 109-slice LIDC-IDRI series. The coronal view uses the DICOM physical spacing. This example has no per-case accuracy score and does not establish nodule-detection performance.
Retinal exudate research view
A research model score map and magnified region on official test image IDRiD_81. These are model scores rather than calibrated clinical probabilities.
Whole-slide candidate localisation research
Retained coarse-grid model scores on CAMELYON17 patient_025_node_0, with a magnified candidate region. The grid is not a precise lesion boundary and the candidate remains research only.
Source data: UPENN-GBM · The Cancer Imaging Archive · Bakas et al., 2022 · UPENN-GBM · dataset DOI · CC BY 4.0. The six MRI views use the retained prediction for one case, UPENN-GBM-00330_11; they are not six independent cases.
CT data: LIDC-IDRI · The Cancer Imaging Archive · Dataset DOI. The current figure uses the LIDC-IDRI-0287 series: 2.5 mm between slices and 0.78125 mm in-plane spacing.
Retinal data: IDRiD · CC BY 4.0 · Porwal et al., 2018 · Data descriptor · Challenge report
Pathology data: CAMELYON17 · CC0 · Bandi et al. · CAMELYON17 challenge
Clinical infrastructure
Built around the way medical data moves.
The wider Nexus platform includes interoperability and operational components, developed for integration with clinical systems and deployment-specific workflows.
FHIR · HL7 · DICOM · structured reports
- Clinical standards
- FHIR and HL7 exchange components, DICOM handling and structured reporting preserve the relationship between clinical records and examination data.
- Recoverable operation
- Persistent records, delivery receipts, retries and backup-and-restore components support continuity when a service or connection is interrupted.
- Practical access
- Browser-based use, bilingual interfaces, resumable transfers and independent workspaces address the everyday requirements of distributed clinical teams.
A second product line
Nexus Agriculture
From field observations, to practical guidance.
An agricultural AI assistant for farmers and extension workers. Share plant photos and field conditions, discuss possible causes and management options, and keep the supporting records together. Saved records can remain available offline; new AI analyses and live data require a connection.
Recorded agricultural experiments
98.4%
Rice image classification
122/124 public Sierra Leone images · task-specific
99.2%
Cassava leaf research experiment
119/120 recorded result · not a released specialist
100%
Cassava root · multiple views
10/10 roots in one small experiment
100%
Whitefly presence recognition
120/120 positive images · no negative-case estimate
91.1%
Weed species research
123/135 images · specialist not enabled online
92.4%
Maize research · top two
133/144 images · top-one accuracy 79.2%
98%
Agronomy QA · any coverage
47 complete + 51 partial answers out of 100
47%
Agronomy QA · complete coverage
47/100 complete · 51 partial · 2 uncovered
These results come from separate experiments and include comparisons of prompts and trained models on reused cohorts. They do not form one untouched blind test or measure performance across all farms. Local specialists remain research assets; the online product currently uses hosted inference.
- Photo diagnosis
- Upload a plant photo, describe the field conditions and discuss possible causes, additional checks and management options.
- Guided follow-up shots
- When one photo is not enough, the assistant names the missing view — whole plant, field context or root cross-section — and guides the follow-up shot.
- Expert advisory
- Discuss agronomic questions and management options. Where relevant, ask about timing and application rates in the context of local products, conditions and guidance.
- Field records
- Structured observations, diagnoses and treatments with a complete evidence chain, exportable for reporting.
- All 14 domains
- Crops, plant health, livestock, aquaculture, weather, markets, machinery, post-harvest and more — 14 agricultural domains on one platform.
- Private workspaces
- Conversations, files and records are isolated per account, with a choice of English or Chinese.
Agricultural engineering
AI shaped by the field task.
Recorded experiments compare hosted reasoning, local classifiers, crop-specific instructions and multiple views. Their results and transfer failures guide further development; research methods are distinguished from features enabled online.
- Task-shaped routing
- The online photo workflow uses hosted inference. Separate research compares it with local crop classifiers and records where results fail to transfer. Those local specialists are not currently active in production.
- Per-crop instruction engineering
- Crop-specific instructions are compared within recorded model and prompt versions. Improvements on a reused cohort remain development evidence and need independent validation before broader claims.
- Guided follow-up shots
- The assistant can request missing views such as the whole plant, field context or a root cross-section. A ten-root experiment reached 10/10 using multiple views; this small result is not a general detection rate.
- Public-archive specialists
- Public-image experiments produced retained weed and maize classifier weights. The final maize evaluation records 79.2% top-one and 92.4% top-two accuracy on 144 images. The 99.2% cassava result remains a historical experiment, not a released model.
- Sealed-evidence discipline
- Original responses, evaluation settings, model versions and negative results are retained. Comparisons and retests are identified separately from independent evaluation.
Field inputs
Photo · question · voiceTask routing
Question context · available toolsHosted inference
Current online analysisFollow-up & aggregation
Missing-view guidance · multi-view seriesRecords & reporting
Evidence-chain field records
Data and delivery
Connected to real data and field reality.
The agriculture line runs on the same operational spine as the medical platform: real data tools, offline-first delivery and per-account isolation, exercised in real browser outage and recovery cycles.
WFP prices · MET/NASA weather · Sentinel · AquaCrop
- Real data tools
- Live market prices from WFP feeds, MET and NASA weather grids, Sentinel satellite pixels and AquaCrop scenarios are called natively when a question needs them — with sources retained.
- Offline-first operation
- Previously saved records can remain readable offline within the authenticated offline-access window, and pending record edits sync on reconnect. New AI answers and live data need connectivity.
- Built for extension work
- Structured field records carry an exportable evidence chain for programme reporting, and the pilot package is designed for extension agents covering thousands of smallholders.
Company
nexus
Nexus develops medical software and specialist research systems. Our work spans clinical conversations, medical-data engineering, specialist image analysis and clinical interoperability. Shared evidence structures and task-specific evaluation let this work accumulate within one platform. A dedicated agriculture line applies the same engineering discipline to crop diagnosis and advisory work.
Contact Nexus- Contactcontact@realnexusgroup.com
- Businessbusiness@realnexusgroup.com
- Careershr@realnexusgroup.com
- Generalinfo@realnexusgroup.com