01
Discover
Find and rank source candidates.
Create datasets that can be measured.
Alys turns a research prompt into local-first dataset artifacts: discovered, verified, deduplicated, structured, benchmarked, and ready for model training.
/dashboard/generate
>
[SRC]ranked 38 source candidates
[CHK]resolved 7 contradiction clusters
[DED]reduced duplicate records by 24%
[EVAL]schema, diversity, utility checks passed
Local-first runtimeJSONL / CSV / RAG / Instruction
PipelineDiscover/Verify/Dedupe/Score/Export
Why Alys
Not a scraper. A dataset engine.
Firecrawl extracts pages. Alys creates evaluated datasets by orchestrating research, verification, synthesis, and export inside one local runtime.
02
Verify
Compare claims and resolve contradictions.
03
Structure
Generate training-ready records.
04
Evaluate
Score schema, diversity, utility, and leakage risk.
Install Alys
Start from your terminal.
Run Alys with npx.
npx alys-akusa
Evaluation
Benchmarks are part of the product.
Statistical fidelity
KS, TVD, Wasserstein, and correlation preservation.
ML utility
TSTR-style downstream model performance checks.
Diversity
Duplicate rate, entropy, redundancy, and mode-collapse signals.
Privacy posture
Nearest-neighbor leakage and membership-inference resistance.
Agentic pipeline
Discovery, extraction, debate, curation, and evaluation are separate modules.
Provider-managed
Users should not bring model keys. Alys owns quality modes, credits, and limits.
Export-native
JSONL, CSV, instruction records, RAG chunks, manifests, and benchmark reports.
Use cases
Instruction tuning datasets
RAG-ready corpora
Structured Q&A generation
Domain research datasets
Sales and support simulations
Evaluation and benchmark sets