# Serokell

> Serokell is a reviewed AI agency in Tallinn, Estonia, listed in the AIAgencies.eu register and last assessed 2026-09-24.

Canonical page: https://www.aiagencies.eu/agency/serokell
Generated: 2026-09-28 · Curated by Gabor Kiss · No paid placements — inclusion, score, and rank are never sold.

## Facts
- Location: Tallinn, Estonia
- Website: https://serokell.io
- Hourly rate: €45-85/hr (directory-listed)
- Team size: 50-249
- Founded: 2015
- Services: AI Development, AI Consulting, AI Automation
- Industries: FinTech, Manufacturing, SaaS & B2B, E-commerce
- LinkedIn: https://www.linkedin.com/company/serokell/

## About
Serokell is a Tallinn-founded software engineering firm with a Paris office, known for Haskell and blockchain work, with a machine-learning practice in predictive maintenance, NLP and recommender systems.

## Documented clients
Documented engagements: Omega Media, InterPop, Biocad, and Tezos Foundation. All 4 carry a source URL on file.

## Curator assessment
Reviewed 2026-09-24 by Gabor Kiss against published evidence.

Five of the 25 case studies Serokell publishes are machine-learning work, and three carry real technical narrative: a locomotive time-to-failure model (XGBoost with a Weibull layer) at F1 above 0.8 for 36-hour forecasts, an NLP parser mapping plain-English requests to 400+ ad segments at over 89% accuracy with a stated 35% cut in campaign cost, and a product-matching and recommender update for a marketplace validated by A/B test.

All five ML clients are anonymous, and the football-analytics case reads loosely—tactics modeled with an LLM on play data drawn from 'PubMed, ProQuest, Elsevier'—so ask for the engineering write-up behind it.

The named clients and most of the team page are Haskell and blockchain work (Tezos, Cardano, Omega Media's Lightning exchange) with one named AI team lead, so ML is one practice inside a functional-programming firm.

Best for buyers who want classical ML—forecasting, matching, text classification—built by engineers who also own the backend and infrastructure, not for buyers looking for an LLM product specialist.

### Key strengths
- ML write-ups publish method and data limits: up to 16% missing timestamps and gaps of up to three weeks in the locomotive data, and why supervised methods were dropped for sparse product categories
- The NLP case states the model moved into production, reporting over 89% accuracy, a 35% drop in campaign cost and twice the lead flow
- A long, verifiable engineering record: founded 2015, a named team page, sponsored work on the Glasgow Haskell Compiler and a research lab run with ITMO University

### Good to know
- Every ML client is anonymous—ask for a speakable reference on the engagement closest to yours
- One named AI team lead against a team page of Haskell and blockchain engineers—ask who on the ML side would staff your project and how many are dedicated to it
- The directories that list Serokell disagree on its hourly rate—confirm which of its fixed-price, time-and-materials or dedicated-team models applies to ML work

---
Source: AIAgencies.eu — the curated register of European AI agencies. Rate provenance: "agency-confirmed" = disclosed via submission or claimed profile; "directory-listed" = Clutch or two agreeing B2B directories (lone unverified sources are never written). Full method: https://www.aiagencies.eu/methodology · Rate index: https://www.aiagencies.eu/rates