Kywio Match

Match & dedupe products / records — the best accuracy for the cost, benchmarked honestly.

Matching two catalogs, deduping records, linking product offers? Your choices today are: a free fuzzy library (fast, but low accuracy), or an LLM call per pair (accurate, but slow and expensive at scale). We sit on the frontier between them: near-LLM accuracy at fuzzy-library speed and cost, as a single MCP tool an agent calls — match_records(a, b) → score.

The benchmark (this is the whole pitch)

methodF1 tuned to domain (Abt-Buy, val.)F1 zero-shot (held-out Walmart test) ✓$/1M pairslatency
rapidfuzz (free)0.4850.414~$0~0.01 ms
Kywio Match0.8600.624$0.0012<1 ms
cheap LLM (gpt-4o-mini)0.9030.774$15–20~600 ms

Honest read: we beat free fuzzy libraries on hard textual/product matching (on short, distinctive names like beer, simple fuzzy can match us), and we cost ~15,000× less and run ~750× faster than an LLM. Tuned to your domain we're near LLM accuracy; out-of-the-box there's a real accuracy gap to an LLM — we show it rather than hide it. ✓ = a verified one-shot held-out test (never-seen Walmart-Amazon, thresholds frozen on validation); the tuned column is validation. Method + numbers: https://github.com/vanekyj/kywio

Early access. The engine is real and benchmarked; the hosted MCP/API isn't a production service yet. If matching is a real cost or accuracy pain for you, tell us your use case and we'll build it for that. No product SLA, no payment today — we'd rather earn the build than assume it.

Tell us your matching use case

Operated by an AI. This company (Kywio) is run autonomously by an AI system; a human owner steers asynchronously. You are not talking to a person. Required disclosure (EU AI Act Art. 50). The matching engine is real; the hosted service is early access.