Project-level scientific memory · EU-first

Ask your lab's own work, and see the source behind the answer.

SciWay reads your team's papers, protocols, ELN entries, results and decision notes, then answers questions with the exact source behind every claim. When the sources disagree, it shows both. When the answer isn't there, it says so.

"What did this team already try, why did they decide it, and what evidence supports it?"
Open the app → Try it on your own text
No sign-up to try · a synthetic demo lab is built in · hosted in the EU, or run the model on your own hardware.
SciWay answering a question with a grounded, cited answer
0.98 citation precision@1measured on SciFact, in a reproducible open-dataset eval
Every answer cites its source, or refusescalibrated refusal is on by default
EU-hosted or on-premisesyour choice of inference; data can stay entirely on your hardware
Why now

A lab's own record is scattered, and the EU is starting to mandate keeping it.

What a lab tried, why it decided, and what failed lives across ELNs, email, file shares, results tables and decision notes. Most of it is never written down twice. General-purpose AI tools answer over this confidently even when the evidence isn't there, and rarely cite or refuse. Meanwhile Horizon Europe, DFG/NWO and NFDI now tie funding to FAIR data management, which turns a lab's own memory into a budgeted requirement.

How it works

The evidence comes first, then the answer.

Ask a question and SciWay pulls the relevant passages from your project and classifies what it found. You get a cited answer when the evidence supports one, both sides when sources disagree, and a refusal when the project doesn't contain the answer.

Supportedgrounded + cited
A supported answer backed by cited source quotes with provenance

The evidence supports an answer, so you get it with the exact source quotes behind it.

Conflictingboth sides shown
Conflicting evidence surfaced, showing two sources that disagree on a value

Two sources disagree on a value. Both are flagged, each with its provenance.

Refusedevidence missing
A calibrated refusal showing what was considered but found insufficient

The evidence is missing, so it refuses and shows you what it considered.

See it yourself

Paste a paragraph and ask it something.

SciWay's deterministic grounding runs in your browser: no sign-up, no API key, nothing leaves this page. The answer is a verbatim sentence from your text. Ask something it doesn't cover and it will refuse.

Try: "what temperature was the stability study?" (answered) vs "what protein concentration was used?" (refused). Synthetic / non-PII content only.
Your grounded answer — with the cited source sentence — will appear here.
Inside the product

One workspace for a project's whole record.

Bring your own documents (PDF, DOCX, XLSX, CSV, MD, TXT, JSON or pasted text), or open the one-click sample project to explore the synthetic demo lab. Trace any claim to its source location, map concepts in the knowledge graph, share a project with your team, and export FAIR metadata.

SciWay project dashboard summarizing a lab's ingested artifacts
SciWay interactive knowledge-graph concept map of a project's corpus
SciWay FAIR export panel: DataCite, Dublin Core and RO-Crate metadata
Capabilities

A trust layer over your corpus.

Every capability keeps the same contract: a claim is only made if a source backs it, and that source is always one click away.

Grounded answers with provenance

Every claim carries its source system, title, page / row / line, exact quote and link. Inspect the evidence behind any sentence.

Calibrated refusal

When the evidence isn't there, SciWay abstains and shows what it considered but found insufficient. On by default.

Conflict surfacing

When sources disagree on a value, both are flagged with their provenance, so a decision can rest on the full picture.

Semantic retrievalLIVE

Embeddings-based search finds the right passage even when the words don't match. A similarity floor turns weak matches into refusals, and you can set how strict that floor is per project.

Typed-relation knowledge graphLIVE

Edges are typed (supports / contradicts / decided / uses-protocol) and carry their own provenance, and entity mentions are linked across documents. Click an edge to see the two chunks behind it.

Cross-project precedent searchLIVE

"Has this team ever tried X before?" A permission-aware search across every project you own or share, under the same citation and refusal contract. Literature search can't answer this; it's a question about your own record.

Scientific tasksLIVE

Assign a follow-up to a teammate with a due date, straight from an answer or a document. They get notified on assignment.

Shared projectsLIVE

Invite your team into a project with owner / editor / viewer roles. The questions you ask and the grounded answers you keep accumulate into a shared record.

Works from your AI assistantLIVE

SciWay runs as an MCP server, so Claude, Cursor or any MCP client can ask your lab's record and gets the same cited answer, or the same refusal. A refusal matters more to an assistant than to a person: it stops filling the gap from its own training data.

Citable evidence export

Export a project's documents, saved Q&A and provenance as a citable evidence log (Markdown / JSON), plus FAIR metadata (DataCite, Dublin Core, RO-Crate).

EU-hosted or on-premises inferenceLIVE

Run answers on an EU-resident model (with PII minimization), on a local model on your own hardware, or with no LLM at all. In that last mode the deterministic engine returns a verbatim cited sentence and nothing leaves the machine.

Ways to run it

Try it in a browser, or keep it all on your own hardware.

The same product in three deployment shapes. Pick the one that matches how sensitive your data is.

Demo lab

Anyone, in 30 seconds

A built-in synthetic lab to explore with no sign-up. Ask the golden questions, see the refusal and conflict states, inspect the citations.

Open the demo →

Hosted workspace

Individual labs & teams

Create an account, upload your documents into isolated projects, ask, and invite your team. Runs on EU-resident infrastructure with per-tenant isolation.

Live  Open your workspace →

Self-hosted, isolated lab

Air-gapped & data-sensitive sites

Deploy the whole stack (store, vector search, graph memory and a local LLM) onto one server inside your network. We install it and hand it off. No data leaves the box, and we hold no credentials to the running system.

In build  Talk to us about a pilot →
Roadmap

Where SciWay is going.

What's shipped, what's in build, and what we're still exploring. The provenance and refusal contract holds at every stage.

Shipped

  • Grounded QA with citations, conflict & calibrated refusal
  • Semantic embeddings retrieval with a calibrated similarity floor
  • EU-resident & local LLM-written answers (opt-in; deterministic by default)
  • Bring-your-own multi-file upload + paste, per-project isolation
  • Shared projects with roles; saved Q&A memory; knowledge graph with typed relations (supports / contradicts / decided / uses-protocol edges)
  • Cross-project precedent search: "has this team ever tried X before?"
  • Scientific tasks: assign follow-ups with a due date, get notified on assignment
  • Citable evidence export + FAIR (DataCite / Dublin Core / RO-Crate)
  • Read-only connector adapters for Zotero, Google Drive and Benchling (fixture-tested; live-API validation in pilots)
  • MCP server, so an AI assistant can query a project under the same citation and refusal contract

In build

  • Self-hosted "isolated lab" box: fully offline, on your infrastructure
  • Bulk corpus import for onboarding an existing project
  • More read-side connectors (LabArchives, Zenodo)

Exploring

  • Multi-hop reasoning across silos
  • Institutional / RDM-office deployments
  • Audit-grade evidence logs & provenance graphs
  • Design-partner pilots (German / Dutch life-science labs)
Trust & EU

Checkable, and honest about its stage.

SciWay is a venture-track prototype, and we say so. Inference is deterministic by default. Any external model call is opt-in, key-gated and PII-redacted, and can be replaced by a model on your own hardware. Bundled lab data is fully synthetic, and the demo accepts synthetic, non-PII content only. Compliance features such as GDPR-aware handling, EU hosting and FAIR-aligned export are design intent for pilots, not a certification.

View the open-source repo →

Tell us where your team loses its memory.

Are you a PI, researcher, or RDM / data steward who keeps re-discovering what your team already tried? Your answers steer what we build next.

Open the app → Take the 2-min survey
Get in touch

Pilots, partnerships, questions

We're talking to academic and translational life-science labs, and their institutions, about design-partner pilots. Hosted or self-hosted.

Email · hello@sciway.software

See it

Open the app

A built-in synthetic demo lab to explore with no sign-up, plus bring-your-own projects: create an account, upload your docs, invite your team, and ask.

Open the app → View the code