Early accessFaraday · The intelligence of your lab
Your best scientist never leaves the building.
Faraday is an AI co-scientist built into a benchtop instrument, and the brain your lab has been missing. It analyzes your data, reads the full-text literature, checks its own statistics and writes up the methods, on hardware that sits in your lab.
- FP4 on NVIDIA Grace Blackwell
- 1 PFLOPFP4 on NVIDIA Grace Blackwell
- Unified memory
- 128 GBUnified memory
- Token context, on the box
- 256kToken context, on the box
- Runs with no network at all
- SealedRuns with no network at all
Named for the cage
Michael Faraday’s enclosure let nothing through. Neither does this one, unless you say so.
Scientific skills
installed on the box
Workflows
across 22 fields
Scientific databases
UniProt to Materials Project
Specialist agents
reviewing its work
The intelligence of your lab
One co-scientist, wired into the whole lab
Faraday sits at the center of your science: reading what your instruments produce, knowing where your data lives, and working alongside every scientist on the team.
Instruments
Mass-spec runs and instrument exports open natively, and the camera reads a display it cannot plug into.
Lab automation
Opentrons and PyLabRobot skills draft and check the protocol for your next run.
Records and ELNs
Skills for Benchling, LabArchives and protocols.io, plus exports from any ELN or LIMS.
Faraday
On your bench, on your network
Your data
Read-only connectors to Amazon S3, Google Cloud Storage and drives on the bench, every import audited.
The literature
Full-text papers, regulatory documents, clinical trials and 110 scientific databases.
Your tools
MCP servers connect GitHub, reference managers and hundreds of other tools.
Where Faraday is going
The brain of the lab, at the bench and at the desk
- Live links to the instruments, ELNs and LIMS in your building
- The same co-scientist at every scientist's desk, not only at the bench
- One memory of every run your lab has ever done
Want it at your desk today? Faraday runs a privacy-hardened build of K-Dense BYOK, our open-source AI co-scientist, which installs on your own computer in minutes.
How it works
From a question to a methods section
Ask the way you would ask a colleague. Faraday runs the analysis in real code, has it checked, and writes down exactly what it did.
- 01
Ask in plain language
Type, tap or dictate. Speech is transcribed on the box in 25 languages, and the articulating camera puts a notebook page, a reagent label or an instrument readout straight into the conversation.
- 02
It works in real code, on your files
Python with pandas, SciPy, statsmodels and the rest of the scientific stack, run against your data. It is built to ask before it assumes, and never to substitute placeholder data or overwrite a source file. Redirect it mid-run without starting over.
- 03
It checks itself before you see it
Specialist agents audit the statistics, the methodology and the references. A citation that cannot be verified is marked unverifiable, never waved through.
- 04
It writes it up as it goes
Every prompt becomes a dated lab notebook entry: the data touched, the code run, the figures made and the sources read, in methods-style prose composed on the box. Export it to Jupyter or Markdown.
You ask
“Run differential expression on counts.csv, treated versus control, and plot a volcano. Flag anything that disagrees with the 2019 result.”
Entry 014
Aug 20, 2026 · 14:22
Recorded · 4m 12s
- Loaded counts.csv18,412 rows
- Ran differential expressionDESeq2
- Wrote volcano.pngfigures/
- Delegated statistical-reviewerpassed
- Consulted frontier expertyou approved
- 9 routine steps folded
Fig. 1The notebook entry that comes back, written while the run happens. Illustrative, synthetic data.
What it does
The four jobs that fill a research week
Each one ends in something you can hand to a colleague or put in a report, not a chat transcript.
Data analysis
A notebook that reruns
Differential expression, survival models, dose-response fits, single-cell pipelines. Every number in the write-up traces back to the code that produced it.
Hypothesis generation
Ranked, with evidence
Testable, falsifiable hypotheses grounded in your data and the literature, each argued with the evidence for and against it.
Deep research
A report with citations
A specific question answered in depth, with a citation checker confirming each reference exists and says what the report claims it says.
Literature review
Full text, not abstracts
Full-text search across bioRxiv, medRxiv, arXiv and PubMed Central, plus regulatory documents and clinical trials, synthesized with proper citations.
Built-in review
A review panel on every run
Faraday delegates to 21 specialist agents: one at a time, in parallel, or in a chain where one's findings feed the next. Edit any of them, or add your own.
Code and computation
statistical-reviewer
Test choice, assumptions, power, multiplicity
code-reviewer
Correctness bugs and numerical pitfalls
data-validator
Schema, missingness, outliers, duplicates
reproducibility-auditor
Seeds, versions and a clean rerun
ml-auditor
Leakage, splits, baselines, evaluation
math-checker
Derivations, units, dimensional consistency
simulation-reviewer
Discretization, convergence, stability
pipeline-engineer
Pipelines that run end to end
data-visualizer
Publication-quality figures
Literature
literature-researcher
Surveys and synthesizes prior work
citation-checker
Each reference exists and supports its claim
fact-checker
Claims against authoritative sources
methodology-reviewer
Threats to validity in the design
peer-reviewer
A full, adversarial journal-style review
Writing
manuscript-editor
Clarity, structure, precision
abstract-writer
Abstracts, summaries, lay versions
ethics-reviewer
Research ethics, privacy, dual use
Study design
hypothesis-generator
Testable, falsifiable hypotheses
experiment-designer
Controls, randomization, sample size
protocol-writer
Step-by-step protocols and SOPs
results-interpreter
Cautious reading, alternative explanations
Honest by instruction
The citation checker must call a reference it cannot confirm unverifiable, never fine. The results interpreter has to offer the alternative explanations before the conclusion.
Scientific depth
Fluent in your field on day one
160+ scientific skills and 326 ready-made workflows arrive installed, with 110 scientific databases a query away, so the box knows your tools before it ever sees your data.
160+
Scientific skills
The libraries your field already uses, with the know-how to use them well. They load on their own, or type / to call one by name.
- scanpy
- pydeseq2
- scvi-tools
- RDKit
- DeepChem
- PyTDC
- ESM
- pyOpenMS
- matchms
- pymatgen
- Qiskit
- Astropy
- PyMC
- statsmodels
- PathML
- pydicom
- NeuroKit2
- Opentrons
Plus NVIDIA BioNeMo
- Boltz-2
- OpenFold3
- RFdiffusion
- ProteinMPNN
- DiffDock
- Evo 2
- GenMol
- MolMIM
326
Workflows
Ready-made protocols for the analyses labs run most. Pick one, fill in the blanks, go. Save your own for the team.
- Single-Cell RNA-seq Pipeline
- Perturbation / CRISPR Screen Analysis
- Retrosynthesis Planning
- ADMET Prediction
- Molecular Docking
- Neuropixels Data Processing
- Digital Pathology Analysis
- Battery Material Analysis
- Gravitational Wave Analysis
- Species Distribution Model
- Reproduce an Analysis
- Write a Specific Aim
110
Scientific databases
Sequence, structure, chemistry, clinical, materials and earth data, queried directly, with the query itself screened by the gate.
- UniProt
- RCSB PDB
- ChEMBL
- PubChem
- Ensembl
- NCBI Entrez
- gnomAD
- GTEx
- KEGG
- Reactome
- STRING
- ClinicalTrials.gov
- openFDA
- Materials Project
- NIST WebBook
- Europe PMC
- OpenAlex
- GBIF
- PhysioNet
- Copernicus CDS
Workflows across 22 fields, including
- Genomics and transcriptomics
- Cell biology and single-cell
- Proteomics and structural biology
- Chemistry
- Drug discovery and pharmacology
- Clinical and health sciences
- Neuroscience
- Materials science
- Physics and quantum computing
- Astronomy and space science
- Ecology and environmental science
- Mathematics and modeling
- Engineering and simulation
- Machine learning
- Social sciences
- Economics and finance
BioNeMo skills call NVIDIA-hosted models and are enabled with your own NGC key.
Native viewers
See your science, not just its files
Structures, spectra and alignments open natively on the workspace screen. Every viewer renders on the box, which is why they still work sealed.
Structure
complex_ab.cif
Mass spec
sample_04.mzML
Alignment
orthologs.a3m
Fig. 2Three of the on-box viewers: a protein structure, a mass-spectrometry run and a multiple-sequence alignment. Illustrative renderings.
Molecular structures
PDB · mmCIF · SDF · MOL2 · XYZ · GRO · PDBQT
Chemistry
SMILES · 2D depiction
Sequences and alignments
FASTA · FASTQ · Clustal · Stockholm · A3M
Genomics
VCF · BED · GFF · GTF · SAM · AnnData .h5ad
Mass spectrometry
mzML · mzXML · MGF · imzML imaging
Documents
PDF with OCR · Word · PowerPoint · Excel · LaTeX · Jupyter
Sovereignty
Your data stays. Every request out is screened.
Faraday does its work on the box, and everything it sends out passes through a gate you can read. When it wants a frontier model's opinion, the request is sanitized, shown to you, and sent only when you approve it.
Outbound payload
Awaiting your approval
Does the vc-PAB linker on [COMPOUND_1] explain the loss of serum stability we measured at DAR 4? [PERSON_1] ran the plates, copy her at [EMAIL_1].
Sanitizer fails closed. If a layer errors, the payload is blocked.
3 spans masked
- KD-7781[COMPOUND_1]Site terms
- Sofia Marchetti[PERSON_1]On-box review
- s.marchetti@northfield-bio.com[EMAIL_1]Pattern
Fig. 3An outbound request, masked by all three layers and waiting for approval.
Three layers, in order
- 01
Patterns
Emails, keys, record numbers, lot codes, InChI and SMILES strings, sequence runs.
- 02
Your vocabulary
Compound codes, targets, cell lines and project names. Your list always wins.
- 03
On-box clean-room review
The local model reads only the outbound text, never the conversation, and can veto it.
It fails closed: if any layer errors, nothing is sent. Approve the masked text unchanged and exactly that text leaves, byte for byte, as a single message with no system prompt and no chat history. Edit it and the sanitizer runs again, so a block can never be approved through.
Every way out, all of them sanitized
Expert questions
Sanitized text to a frontier model, sent only after you approve the exact payload.
Images
Pixels cannot be masked, so every image needs your explicit approval. Every time.
Web searches
Sanitized first. Clean queries run. A query that carries anything proprietary waits for you to edit it.
Database queries
Queries to the scientific databases are sanitized, then run without a per-request approval.
BioNeMo models
Calls to NVIDIA-hosted models, on your own NGC key, are sanitized and run without a per-request approval.
Connected tools
MCP tool calls are sanitized and run without a per-request approval.
Seal the box and every one of these paths closes.
A dial, not a switch
Set how much can leave, per session, from freely connected to fully sealed. Every change is on the record.
Egress
Seal the box: local models only, no web, nothing out.
Sanitize
Every outbound request masked before it goes.
Preview
You approve each expert request. Flagged searches wait for you.
Tightening is always allowed. Loosening the gate takes an administrator, and the role comes from the operating system on the box rather than the screen. Sealed, Faraday makes no outside connections of its own and still runs voice, document parsing, every viewer and the notebook.
Provenance
Every result arrives with its paper trail
Faraday treats an answer the way you treat a measurement: traceable to its inputs, its method and the instrument that produced it.
01
A lab notebook that writes itself
One dated, numbered entry per prompt, with the data, code, figures, citations and gate crossings captured as they happen. Add margin notes and flags, then export to Jupyter with one code cell per step.
02
A hash-chained audit log
Everything that crosses the boundary, out and back in, is written to an append-only SHA-256 chain before it is sent. An independent verifier, sharing no code with the product, re-checks every hash.
03
Per-person accountability
Each record carries the operator's own account on the box, so the trail is per scientist without a separate identity system to administer.
04
Runs that end honestly
Every reply closes as Complete, Stopped or Failed, with its duration and step count, so a quiet minute never passes for a finished analysis.
The instrument
Operated, not installed
Results and the record on the large screen, the conversation on the small one. No desktop and no terminal for lab users, just the instrument.
Fig. 4The instrument, from the front right.
- 1
15.6" workspace touchscreen
Files, data previews, the scientific viewers and the live lab notebook.
- 2
11.6" chat touchscreen
The conversation, the composer and every approval the gate asks for.
- 3
Twin side fans
Cooling for long, sustained analyses on the Grace Blackwell chip.
- 4
Optional keyboard
With a built-in touchpad, for editing code and notebooks.
Touch first
Every control sized for a fingertip, and a text-size control from 80 to 140 percent for reading at arm's length.
Voice, on the box
Dictation is transcribed locally by NVIDIA Parakeet in 25 languages. The audio never leaves.
A camera for the bench
Digitize a notebook page, identify a reagent, read an instrument display or document a setup in one capture.
Always warm
The model stays loaded across logouts, reboots and updates, so the instrument is ready when you walk up to it.
Datasheet
Specifications
- Compute
- NVIDIA GB10 Grace Blackwell superchip
- AI performance
- Up to 1 PFLOP FP4
- Memory
- 128 GB unified LPDDR5x
- Co-scientist model
- Gemma 4
- Model options
- NVIDIA Nemotron, other local models
- On-box context
- 256k tokens
- Also on the box
- Speech-to-text, OCR, document parsing
- Frontier expert
- Optional, only through the gate
- Displays
- 15.6" workspace and 11.6" chat touchscreens
- Capture
- Articulating camera, microphone
- Input
- Touch and voice, optional keyboard
- Storage
- USB or Thunderbolt, typically 2 to 8 TB
- Connectors
- Amazon S3, Google Cloud Storage, local drives
- Accounts
- Per-user, Admin and User roles
- Software
- Privacy-hardened build of K-Dense BYOK
- Interface
- Locked kiosk across both screens
- Siting
- Benchtop, standard wall power
Powered by Gemma 4, running entirely on the box. If your group has standardized on a different model, Faraday also runs NVIDIA Nemotron and other local models.
Delivery
Results in week one
We do the siting, the integration and the training. You get a result before the invoice is due.
Week 0
We site it
Standard bench, standard wall power. No rack, no room works.
Day 1
Your data, read-only
Imported through the connectors, and your own vocabulary loaded into the gate.
Same day
First analysis
Your team asks in plain language. The notebook starts writing itself.
Week 1
Measured, not asserted
Your pre-registered baseline against what actually happened on the bench.
Next-business-day response, same day if the box is down, a weekly check-in and a dedicated channel to your team. Updates install from Settings with a single button, and both screens come back exactly where they were.
Compared
The four ways this problem gets solved
Faraday does not win every row. It wins the two that usually decide it.
| Dimension | Faraday | Hire for it | Internal queue | AI in your cloud tenant |
|---|---|---|---|---|
| Time to first result | Same day | Months to recruit | Weeks in the backlog | Days to provision |
| Breadth of domains | 22 fields, 160+ skills | One person's expertise | Whatever the team knows | Broad |
| Where your data sits | Your building | Your building | Your building | A hyperscaler region |
| What leaves during a run | Only what you approved | Nothing | Nothing | Your data |
| Several questions at once | Parallel chats, in the background | One at a time | The queue grows | Elastic |
For your security review
What we claim, and what we do not
Trust in the gate is the product, so we would rather your reviewers hear the boundaries from us.
True today
- The gate masks, previews and logs every expert request, and fails closed if any layer errors.
- The co-scientist, the gate's review and speech-to-text all run on the box. Analysis continues with no network.
- Connectors are human-driven, cloud storage is read-only, and none of them is exposed to the agent as a tool.
- Roles come from the operating system identity on the box and are enforced server-side, never claimed by the client.
- An independent verifier, sharing no code with the product, re-checks every hash in the audit chain.
Stated plainly
- The gate governs what Faraday itself sends. For a fully sealed site, pair it with your network controls, and we will map exactly where that line sits with your team.
- The audit chain is tamper-evident, not tamper-proof.
- Zero data retention at the frontier model is contractual, and we document the terms for your review.
- Architected toward HIPAA, SOC 2 and GxP. Not yet certified against them.
FAQ
Questions scientists ask us
Does our data ever leave the building?
Your data stays on the box, where Faraday does its analysis. Every request Faraday sends out, from web searches and database queries to connected tools, passes through the sovereignty gate, which masks sensitive text and logs every exchange. Questions for a frontier model also wait for your approval, and images always do. Seal the box and Faraday makes no outside connections of its own at all.
Which AI model runs on the box?
Gemma 4 runs locally on the NVIDIA Grace Blackwell chip, and an administrator can switch to NVIDIA Nemotron or another local model. When a question needs a frontier model, Faraday can consult one through the gate, sending only sanitized text you approved.
Can it run fully offline?
Yes. Sealed, Faraday uses local models only, with web access and outside tools removed. Voice dictation, document parsing with OCR, every scientific viewer and the lab notebook keep working, and data comes in on direct-attached drives. Web and full-text literature search need a network, so they pause while the box is sealed.
Is our data used to train anyone's model?
No. Your data stays on hardware you own. The sanitized requests that do go to a frontier expert are sent with logging and training switched off, under zero-data-retention terms we document for your security review.
How do we reproduce an analysis?
Every run leaves its scripts, outputs and a dated notebook entry in the project. Export the notebook to Jupyter and each step becomes its own code cell, ready to rerun.
What file formats does it understand?
Tables, PDFs and Office documents (with OCR for scans), molecular structures such as PDB, mmCIF, SDF and MOL2, SMILES, sequences and alignments in FASTA, FASTQ, Clustal and Stockholm, genomic tracks such as VCF, BED and GFF, AnnData, mass-spectrometry runs in mzML, mzXML, MGF and imzML, LaTeX and Jupyter notebooks.
How many people can use one unit?
Faraday is built for a small team. Each person gets their own account and their own audit trail, with Admin and User roles. One person drives the instrument at a time, and can run several chats in parallel.
Is it validated for GxP, HIPAA or SOC 2?
Faraday is architected toward HIPAA, SOC 2 and GxP, including 21 CFR Part 11, with hash-chained audit records, role-based access and controlled updates. It is not yet certified against them, and we will walk your quality and security teams through exactly where it stands.
How is Faraday different from K-Dense Web and K-Dense BYOK?
Mostly in where the work runs and what can leave. Faraday is an instrument on your bench: its co-scientist model runs on the box, every outbound request is sanitized by the sovereignty gate, and questions for a frontier model wait for your approval. K-Dense Web is our hosted platform: it runs in K-Dense's cloud on frontier models we choose, and carries a research task through to finished deliverables once you approve its plan. K-Dense BYOK is free and open source: it runs on your own computer with the models and API keys you pick, and Faraday runs a privacy-hardened build of it. K-Dense Enterprise deploys K-Dense Web in a dedicated cloud, your own cloud or your data center. The comparison table on this page sets out every difference.
Is Faraday available now?
Faraday is in early access. Request a demo, and we will show you the instrument running on work like yours and talk through timing for your lab. Pricing is quoted on request.
Can we try it before we buy?
Yes. Faraday runs a custom build of K-Dense BYOK, our free, open-source AI co-scientist. Install BYOK on your own computer to try the same research agent, skills and workflows today. Faraday's build puts privacy and data security first, adding on-box models, the sovereignty gate, the hash-chained record and the locked instrument around it.
Choosing a product
Faraday, K‑Dense Web, BYOK or Enterprise?
Customers ask us how these differ. Most of the answer is where your work runs and what is allowed to leave.
| Compared on | Faraday | K‑Dense Web | K‑Dense BYOK | K‑Dense Enterprise |
|---|---|---|---|---|
| Basics | ||||
| What it is | An AI co-scientist built into a benchtop instrument | A hosted AI research agent in your browser | A free, open-source app on your own computer | K‑Dense Web, deployed for your organization |
| Best for | Labs whose data cannot leave the building | Finished deliverables with nothing to install | Hands-on control over models, keys and cost | Organizations that need their own cloud, SSO and SLAs |
| Price and license | Quote on request | Instant answers free, then pay as you go. Plus $199/mo, Team $499/mo | Free, MIT license. You pay your model provider | Custom quote |
| Getting started | Early access. Request a demo, and we site it for you | Sign in at app.k-dense.ai. Nothing to install | Install from GitHub in about five minutes | Talk to us. We scope, deploy and onboard |
| Data and security | ||||
| Where it runs | On the instrument, on your bench | K‑Dense's managed cloud | Your own computer | A dedicated K‑Dense cloud, your AWS, GCP or Azure, or your data center |
| Where your data goes | Stays on the box. Every outbound request is sanitized by the gate, and expert questions wait for your approval | Uploaded to K‑Dense's cloud and sent to frontier model providers. Encrypted, never used for training | Stays in folders on your computer. Each request goes to the model provider you pick | Stays inside your own cloud or data center, with models on endpoints you control |
| Works offline | Yes, sealed. Web and literature search pause | No | With local models. Web search needs a connection | Air-gapped installs by arrangement |
| Compliance | Architected toward HIPAA, SOC 2 and GxP. Not yet certified | Not SOC 2 audited. No SOC 2 or HIPAA claim | No certification. Runs under your own IT controls | Runs in the environment your security team governs. Not SOC 2 audited |
| Capability | ||||
| AI models | Gemma 4 on the box, with NVIDIA Nemotron and other local options. A frontier expert only through the gate | Frontier models, chosen by K‑Dense | Your choice: OpenRouter, direct API keys, supported subscriptions or local models | Frontier models chosen by K‑Dense, or models on your own cloud endpoints |
| Compute | NVIDIA GB10 Grace Blackwell with 128 GB unified memory, on the box | K‑Dense's managed cloud | Your computer, with optional Modal cloud CPUs and GPUs | Dedicated resources on K‑Dense's cloud or yours |
| Skills, workflows, databases | Scientific skills, workflows and database access, installed on the box | A large library of databases and on-demand tools, hosted for you | The open-source skills library, workflow templates and database guides | Everything in K‑Dense Web, plus connectors to your internal data |
| What you get back | A dated notebook entry per prompt, with figures and reports. Export to Jupyter or Markdown | Manuscripts, slides, posters, PDF reports and figures, with the code and data | Scripts, figures, reports, a living lab notebook and a LaTeX editor | Everything in K‑Dense Web, plus custom reports |
| How you work | ||||
| Working style | Interactive at the bench: touch, voice and camera. Redirect it mid-run | Approve a plan, then it runs to completion on its own | Interactive: asks as it goes, steer it mid-run, parallel chats | Same as K‑Dense Web |
| Team use | A small team: per-person accounts, Admin and User roles, one operator at a time | Team plan: unlimited seats and a shared credit pool. Share sessions by link | One person, on their own computer | SSO/SAML, role-based access, team workspaces and usage analytics |
| Support | Next-business-day response, same day if the box is down, weekly check-ins and a dedicated channel | Email and office hours. Priority for Plus and Team | GitHub issues, docs, tutorial videos and office hours. Community-supported, no response-time guarantee | Priority support, a dedicated success manager and SLAs |
| Audit trail | Hash-chained log of everything that crosses the boundary, per-person records and a self-writing notebook | Full session history, with each run's code and data downloadable. Formal audit logs are an Enterprise feature | Living lab notebook of hypotheses, methods, code and decisions, plus cost per run | Audit logging across agent activity |
| Next step | Request a demo | Explore K‑Dense Web | Try BYOK on GitHub | Explore Enterprise |
Scroll sideways to see every product.
Faraday runs a privacy-hardened build of K‑Dense BYOK, so you can try the same agent on your own computer today.
Put an AI co-scientist on your bench
Tell us what your group is working on, and we will show you the instrument running on work like yours.