With 26 specialized AI agents, AiLabrix accelerates the path from clinical dataset to review-ready report, automating analysis, validation and documentation.
Ingestion templates, QC gates, statistics and reporting tuned per assay — with deep-dive playbooks for each one.
From CSV upload to signed report: every step is tracked, verifiable, and designed for teams that need evidence defensible in audit.
Drag-drop CSV, TSV, Excel. Automatic detection across 12 assay types (PCR, immunoassay, flow cytometry, mass spectrometry, sequencing, imaging, drug screening, microbiology, cell biology, cell viability, clinical chemistry, metabolomics). Optional <filename>.lims.json sidecar as ground-truth input to the arbiter.
18-category sanitization (names, addresses, dates, email, SSN, MRN, IP, CF italiano, ...) runs before any LLM touches the data. Two-pass defense: deterministic regex + lookup, then LLM verification on residuals. Audit records column names + categories — never raw values.
Parametric vs non-parametric auto-selected from Shapiro-Wilk + Levene. Hedges' g, rank-biserial, epsilon², odds ratio — all with 1000-iter bootstrap-BCa confidence intervals. Benjamini-Hochberg FDR stratified per family (biomarker discovery / confirmatory / demographic / QC).
Deterministic LangGraph orchestration with a Fase 3 LLM arbiter for ambiguous assay detection. Fires only when keyword + column-semantic confidence < 70 or disagreement. Anti-bias alphabetical prompt, SHA-256 cache, automatic fallback to keyword top-1 on parse error.
Okabe-Ito daltonic-safe palette, Inter Display + JetBrains Mono, 300 dpi PNG, editable SVG (text stays text), vector PDF with TrueType-embedded fonts. Every SVG carries provenance (pipeline version, git SHA, dataset hash, model ID, seed, timestamp) in a structured metadata block.
Append-only action log (login, gate, config). Per-LLM-call audit (model, tokens, cost USD, latency, outcome) with 90-day retention and GDPR Right-to-be-Forgotten query by dataset hash. Compliance flags per assay: MIQE (PCR), STARD (immunoassay), CLSI EP28 (clinical chemistry), MINSEQE (sequencing).
Automatic enrichment against HMDB (metabolomics), KEGG + Reactome pathways (hypergeometric), STRING PPI networks, MSigDB hallmark GSEA, LOINC clinical reference ranges. Async via Dramatiq queue with exponential-backoff retry and a 30-day disk cache. Fail-soft: API outages never block the report.
Reproducible precision/recall baseline: 60 synthetic datasets (easy / medium / hard variants across 12 assay types, deterministic seeds) + 60 public pointers (GEO, MetaboLights, PRIDE, ImmPort, PhysioNet). Measures top-1 / top-3 accuracy, per-type F1, latency p50/p95/p99, cost USD per call.
Switch the reasoning engine without redeploying: Anthropic Claude, OpenAI GPT, or fully-local Ollama. Set AILABRIX_EU_MODE=1 to disable every cloud LLM and force on-prem inference — zero egress, full GDPR posture. Self-hosted single-tenant or EU-hosted SaaS, your choice.
RUO software with a compliance-ready spine: e-signatures, CAPA, vigilance triggers, PMS dashboard, validation harness with hash-stamped evidence ZIP. Not certified — but the artefacts an auditor expects are already there, generated nightly, and traceable to ISO 13485, IEC 62304, ISO 14971, ISO 15189, IVDR and 21 CFR Part 11.
Every signing event re-authenticates the operator (no session reuse), captures intent + statement, computes a SHA-256 over user_id + artifact + statement + microsecond timestamp, and writes an immutable signatures row plus an audit chain entry. verify_signature_chain() walks the whole table for forensic integrity checks.
open → investigating → action → verifying → closed. Closure REQUIRES a signed approval (uses the e-sig module). Per-CAPA: root cause, action plan with owners + due dates, effectiveness check, signed closure with chain-of-custody to the source incident.
6 event types (job_failure, detection_drift, qa_gate_blocked, clinical_flag, user_complaint, audit_chain_break). 4 trigger rules: R1 critical severity, R2 clinical impact, R3 repeat-rate (≥3 events on same playbook in 14d), R4 audit chain break. MDR/IVDR-style summary export ready for vigilance reporting.
10+ KPIs roll up nightly: jobs/month per phase, failure rate, CAPA throughput + median time-to-close, open vigilance events, signature throughput, audit chain integrity score. CSV export for export to GRC tools. Monthly PMS PDF auto-rendered on the 1st of each month with provenance footer.
Welch's t via scipy.stats, BH FDR via statsmodels, Kaplan-Meier + log-rank via lifelines, PERMANOVA via scikit-bio. Every run captures library versions + computes SHA-256(versions, prompt_hash, seed) — identical inputs guarantee bit-for-bit identical numerical output. 30 golden agreement tests pin 1e-9 tolerance vs reference impls.
21 validation protocols across 17 healthcare R&D phases run nightly on GitHub Actions, with drift detection (new_failures / new_passes / still_failing) and Slack alerts on regression. One-click build_evidence_package.py emits a hash-stamped ZIP with SDLC, RMF, VVR, traceability matrix mapping every protocol to its ISO/IEC clause, 90 days of run records, audit chain snapshot, and library provenance — ready for hand-off to a notified-body consultant.
Compliance posture — read this first: AiLabrix is research-use-only (RUO) software. The features above are compliance-ready by design but the platform is NOT certified for IVD diagnostic use. Certification (CE-IVD / 510(k)) requires a notified-body audit on the customer's organizational SOPs in addition to the technical artefacts AiLabrix ships. We help you build the dossier — you own the submission.
From raw file to signed report. AiLabrix guides you step by step, always showing what's happening under the hood.
Drag CSV or Excel files. AiLabrix detects schema, assay type and suggests the most likely merges.
26 agents run in parallel: profiling, QC, inference, figures, drafting. You see live status on a timeline.
Each gate asks for your OK. Contextual annotations, versioned comments, no emails lost in threads.
Signed PDF, raw + intermediate data bundle, complete analysis log. All with SHA-256 hash.
Most platforms stop at "the AI said so." AiLabrix goes further: every scientific interpretation is peer-reviewed by a second AI, and every LLM-backed agent can run two models in tandem to cross-check outputs before they reach the report. Two mechanisms, always transparent, always auditable.
Inspired by DeepMind's Gemini Deep Think, the scientific interpretation is generated as K parallel candidates (Aletheia role), each tested with balanced prompting — one pass arguing the result is correct, another arguing it is wrong. A tool layer runs inside the loop: code-assisted numeric verification against raw results plus web-grounded search (Brave → Google → DuckDuckGo fallback). An Advisor ranks the candidates and emits a structured verdict; a Colleague role challenges the winner with the strongest alternative hypothesis. Compute is log-scaled by level (L1 → L4).
One AI is smart. Two AIs that disagree and resolve are scientifically stronger. Dual AI lets you pair a primary model with a reviewer model — per agent — and pick one of two validation modes. Every output carries a SHA-256 hash and full provenance (which model wrote what, which revised what) in the report appendix.
Primary writes. Reviewer marks errors. Primary revises. Sequential, deepest quality.
Both models reason concurrently. Primary merges the strongest claims. Fastest, diverse angles.
30 seconds are enough to see a clinical dataset turn into a signable report. No jupyter notebooks, no fragile R scripts, no lost Excel files.
AiLabrix replicates a real R&D laboratory as a software organization. 26 AI and Python agents, grouped in four divisions under a C-suite, pass the dataset along the same chain a human lab would: intake → statistics → science → quality. Each role is specialized, each handoff is logged, nothing skips steps.
The bench team. Receives the raw dataset, detects schema and assay type, fixes missing values, runs exploratory analysis and checks chart readability before anything moves downstream.
The biostatistics desk. Chooses the right test based on distribution and sample size, applies multiple-comparison corrections, computes effect size and 95% CI, and halts if the data aren't powered for inference.
The scientists' room. Turns numbers into a biomedical narrative, stress-tests it with peer review (Deep Think), pulls supporting literature, curates evidence and maps findings to clinical relevance.
The QA office. Runs compliance checks against ISO 15189 / GxP principles, assembles the append-only audit trail, generates the signable PDF and seals every artefact with a SHA-256 hash.
AiLabrix generates figures compliant with scientific journal standards. Every chart is signed with source dataset and tested hypotheses.
Research teams that need defensible, reproducible and fast results, without depending on a single data scientist.
Corporate teams of 3-20 researchers running 10+ experiments per month who need to standardize analysis and reporting across different people.
"We halved the time between end-of-experiment and signed report. And we finally know how every number was obtained."
Those serving different clients on different datasets need an identical, certifiable workflow that reduces human error and validation effort.
"Every client wants the same report but on different data. With AiLabrix the pipeline is the same: only the input changes."
Research groups that must deliver clean analysis to clinical teams and regulatory affairs. Audit trail, versioning, hash on every artifact.
"Regulatory is happy. We're happy because we no longer spend weeks rebuilding 'how we calculated that mean'."
Log in if you already have an account, or drop us a line: we'll reply within 24 hours with an onboarding tailored to your lab.