Watch SANA work: bring data in, analyse, ask the assistant, read the report. Every screen below is SANA itself.
What SANA solves
Labs that test and measure samples run into underlying problems again and again: a result that looks good but is not really true, a file from one instrument that will not open in another program, a number that nobody can check later. Below are ten of these everyday problems and what SANA does about each one.
Tests every model against shuffled copies of your own data and by cross-validation, then says Supported, Use with care or Not supported. If a model does no better than guessing, SANA tells you so plainly.
Reads exports from 88 kinds of instrument, including native SPC, OPUS, OMNIC, ChemStation, Bruker NMR and mzML. SANA Sync sends finished files from the instrument PC.
Routes the data by its shape and your goal to the right engine, explains why in the report and cites the method from its knowledge base.
Resolves overlapping signals into pure spectra and profiles (MCR-ALS, PARAFAC) and reports how many components the data really supports.
Flags unusual samples (T², Q, cross-validated error). You can re-run the analysis without them. The report lists each sample you removed and how many flags would appear by chance.
Flags new samples outside the model's domain, transfers calibrations (SBC, DS, PDS), maintains models (EPO, GLSW) and watches drift in routine QC.
Calibration lines with linearity, LOD/LOQ, range and recovery; reports signed as author, reviewer or approver with a SHA-256; everything kept in History.
Every step can be taken by hand. Any session downloads as a Python script that reproduces it, and an independent R implementation gives a second opinion.
Six ready-made examples you can run with one click, a guide and Ask SANA: say what you need in plain words, approve the plan and SANA does the analysis.
Ask SANA (the AI assistant) only sees file names, column names and SANA's own verdict wording with the numbers taken out. It plans from that, and SANA's engines do the computing.
What SANA can analyse
Seven kinds of analysis cover what lab instruments measure. Choose your instrument, or let SANA work out the right analysis from your data. SANA's own tested programs do all the calculations. The AI assistant only plans the steps and never works out a result.
Predict a property from spectra or signals, validated by cross-validation and against chance.
PLS · PCR · MLR · local PLS
Fit kinetics, decays and dose-response curves, with each parameter's interval, or a plain note when it cannot be pinned down.
model choice · AICc/BIC · runs test
Separate a run's overlapping signals into pure spectra and elution profiles, and estimate how many components the data may support.
MCR-ALS · EFA · ambiguity bands
Decompose a batch of excitation-emission matrices or runs into its components and check the model holds (core consistency).
PARAFAC · PARAFAC2 · Tucker
Map where each material is in an image cube, with a pure spectrum for each, even for cubes larger than memory.
unfold · MCR-ALS · out-of-core PCA
Tell classes or origins apart, with accuracy tested against shuffled labels and a confusion matrix.
PLS-DA · LDA · QDA · kNN · RF
See the structure in samples without labels: score plots, loadings and clusters, with outliers flagged.
PCA · HCA · k-means · T² / QEach with example data, each kept in History with its own report and offered as "next checks" after an analysis.
Calibration line with ICH Q2 figures · method comparison (Bland-Altman, Deming, Passing-Bablok) · chromatogram integration and system suitability · process monitoring (T², Q) · design of experiments · combining instruments
Descriptive statistics · t-tests · ANOVA · correlation matrix · chi-square test
Shewhart, EWMA and CUSUM charts · capability (Cp, Cpk) · precision · GUM measurement uncertainty · robustness (Plackett-Burman)
DD-SIMCA authentication · SVR and GP vs PLS · PLS2 · O-PLS / O-PLS-DA · variable selection (VIP, CARS, GA) · neural network and boosting · data fusion · EPO, GLSW
Spectral library search · band fitting · MS mass accuracy, isotope pattern and formula finder · NMR processing and qNMR · untargeted MS
Thermal kinetics (Kissinger, OFW, KAS) · DSC purity · equivalent-circuit fit · particle-size distribution · mixture and D-optimal designs · desirability
Missing and below-LOD values · batch trajectories (multiway PCA) · chemical image segmentation · cluster analysis
PARAFAC2 · Tucker3 / Tucker2 · kinetic hard-model MCR · Bayesian calibration · metabolomics (volcano) · 1-D and image CNNs (optional)
Why you can trust the result
A result is only worth trusting if it has been checked. SANA checks every result it gives you. It also checks its own code after every update, so mistakes are caught before they reach you.
When the data is not good enough for a reliable answer, SANA says so plainly in the report. This is tested for every kind of analysis.
SANA scrambles the answers in your data and runs the same method again. If your real result is not clearly better than the scrambled one, SANA does not trust it. It also tests each model on samples it has not seen. A more complex method is used only when it is clearly better.
R is a widely used statistics program. SANA can repeat calibrations, curve fits, classifications and 13 lab tools in R, using code that was written separately from SANA's code. R also draws its own figures. SANA then compares the two sets of results and tells you whether they agree.
A separate program tests SANA the way you would use it. It hides the right answers, then checks SANA's results against published work and against data with known answers. It runs after every change.
Each report explains why SANA chose that analysis and method. It cites the knowledge base the method comes from and says what the result cannot tell you.
You can do each step by hand, with SANA's suggestion shown beside it. You can also download a Python script that repeats the whole session, so anyone can run it again.
Each file is opened by a short-lived program with strict limits and no access to passwords. If a file is damaged or harmful, SANA says so plainly and keeps working.
You can sign a report as author, reviewer or approver. The signature records the exact time and a fingerprint of the report. If the report is changed after signing, this shows. SANA's validation summary lists only checks that really exist.
Ask SANA · AI assistant · coming soon
Tell Ask SANA what you want in plain words. The assistant plans the steps and SANA's tested engines do the calculations.
Numbered steps in plain words. Nothing runs yet.
Approve, untick steps or stop. A step you removed can never run.
Each step runs as you, through SANA's own functions and lands in History.
Figures and verdicts are SANA's. The assistant's words are shown apart and labelled as AI-written.
<sana_data> { "file": "F1", "name": "example-nir-spectra.csv", "shape": [120, 701], "columns": ["analyte", "1100", "…"] } { "analysis": "A1", "status": "done", "verdict": "PASSED: R2cv of # sits far above what the same model achieves on # shuffled copies" } </sana_data>
Use SANA's assistant or your own Claude, ChatGPT or Gemini key. In an evaluation of 97 lab tasks with real Claude models, no measured value or sample name reached the model, and no instruction planted in a file was followed.
Ask SANA is switched off while SANA is in testing. Separately, as a testbed, SANA may keep what you upload, and Lavaka may look at it and use it, including with AI models, to improve SANA: see the privacy policy.
Ask SANA (measured)
Real models drove SANA through 97 realistic lab requests, scored by fixed rules from SANA's own records, never by another AI. Model planning is by AI but SANA's engines compute the numbers.
| Model | Requests done fully right | Right plan | Same result on a repeat | Privacy and trap tests passed | Measured values sent to the AI | Cost per request | Time per request |
|---|---|---|---|---|---|---|---|
| Claude Opus 5.5Anthropic | 92 % | 95 % | 92 % | 95 % | 0 | $0.65 | 31 s |
| Claude Sonnet 5.5Anthropic | 88 % | 95 % | 92 % | 90 % | 0 | $0.25 | 22 s |
| Claude Haiku 5.5Anthropic | 78 % | 83 % | 80 % | 90 % | 0 | $0.013 | 18 s |
| GeminiGoogle | Being tested next, on the same tasks and rules. | ||||||
Main evaluation: 449 runs, each request twice per model; privacy and trap tests: 63 runs (prompt injection, impossible and out-of-scope requests). Measured in November 2026, before the fixes the evaluation led to; after them the held-out results rose (Opus 87.8 → 92.7 %, Haiku and Sonnet 81.7 → 84.2 %). Not one measured value or sample name reached any model, and none followed an instruction planted in a file.
Gallery
Five figures for each of SANA's seven analysis types, exactly as SANA draws them in its reports: from its worked examples and from real public datasets. Every reference line and band is named in a legend, colours are safe for colour-blind readers and downloads are 300 dpi or vector. Swipe or use the arrows; open a figure to see it large.
Instruments
Spectroscopy, mass spectrometry, chromatography, imaging, electrochemistry, thermal and physical methods.
In your lab
A small Windows app: pair once, choose the folder and finished files go out over outbound HTTPS only.
SANA reads each file in a sealed process, runs the analysis and gives its verdict.
New files checked against approved, signed model versions, with alerts and drift.
Results back to the instrument PC exactly once, a scoped API and LIMS layouts.
SANA is in testing and free to use. Log in with your Lavaka account and ask for SANA on your account page; we approve each person. As a testbed, SANA may keep what you upload, and Lavaka may use it to improve SANA (privacy policy).