Rohlik Group · Engineering

The stack behind Europe's e-grocer

Real software meets the physical world — warehouse robotics, ML forecasting and an AI-native engineering culture, across 342 active repositories in five markets.

Agentic contributions · 30d
85%of contributions are now agentic
↓ 19%
PR review wait vs 90-day baseline
↑ 38%
merged PRs MoM
196
repos with Claude Code

Proud of the leverage, clear-eyed about the misses — and trusting it: some merges already ship on AI review alone.

342
Active repositories
of 625 scanned · GitLab, June 2026
221
Contributors
across 5 markets
~58
Deploys / day
1,738 prod deploys / 30d · Datadog DORA
8,350
Commits / 30 days
≈ 278 a day · ↑ 10%

Sources: GitLab API scan (June 9, 2026) + Mesmer delivery analytics + Datadog DORA (prod deploys)

What we build with

89 technologies across 7 domains, detected from 625 repositories and curated by the engineers who own them. Starred items are the paved road — our company-wide defaults.

GrowingSteadyNewSunsetting
89 of 89 technologies

Languages

9

Two centres of gravity — Python for ML, data and agents; Java for the transactional core. Python is shown as two estates: ML/data and services.

Backend

14

Spring Boot on the JVM, FastAPI on Python — integration-tested against real databases, resilient by default.

Frontend

12

React everywhere — storefronts, admin tools and the fleet hub, increasingly on Next.js, Vite and Tailwind.

Mobile

6

Kotlin-first Android with Compose and KMP, Swift on iOS — plus on-device ML for quality inspection.

Data & Databases

14

MySQL and Postgres up front, Snowflake behind, RabbitMQ in between — and DuckDB federating across it all.

AI & ML

21

Models in production, agents in the workflow, ML in every forecast — the AI-native layer of the stack.

Infrastructure & DevOps

13

GitOps on GKE — 442 repos ship through GitLab CI into ArgoCD canary rollouts, watched by Datadog.

Where software meets the physical world

Groceries are unforgiving — fresh food, tight slots, real robots. These are the problem spaces the stack exists to solve.

Warehouse automation

AutoStore robotic grids across our fulfillment centers. In-house services bridge the WMS to physical robots — decanting, automated picking, conveyor routing and dispatch. A pick-balancing engine uses mixed-integer programming to decide which SKUs move to the grid.

AutoStoremixed-integer programming

Supply chain optimization

End to end: ML demand forecasting feeds purchase orders feeds supplier delivery reservations. The labour planning engine turns predicted demand curves into optimal shift patterns with linear programming.

ML forecastinglinear programming

Last-mile logistics

Route planning via in-house solver engines, real-time fleet tracking and courier capacity optimization. ML predicts handling times; five markets run different delivery models, from 60-minute express to eco slots.

route optimizationfleet tracking

Payments & fraud

Adyen, PayPal and bank transfers across markets, with real-time fraud scoring built in.

AdyenPayPal

AI-native engineering

MCP is an API surface across 15+ services. Devin, Claude Code and PR Agent are embedded in the daily workflow, and Sandstorm runs agents in isolated sandboxes for operational analysis. 85% of contributions are agentic, and some MRs already merge on AI review alone.

MCPClaude CodeSandstormDevin

Multi-market architecture

Every service is multi-tenant — isolated data per market with a shared codebase. Regional identity servers, market-specific feature flags and tenant-aware routing are infrastructure primitives, not afterthoughts.

CZ · DE · AT · HU · RO

Built by small cells

Product cells own their problem end to end — small enough to move fast, accountable enough to be woken up by what they ship. ML, AI and OR engineers sit inside the cells, not beside them.

3–5
Engineers per cell
“Five people can hold a problem in their heads. Fifteen can't.”
1
North Star metric per cell
Public and non-negotiable, with 2–3 guardrails it must not break
24/7
You build it, you run it
P1 incidents route straight to the owning cell via incident.io
MaiaConversations / dayCSAT must not drop
AdsAds revenue per MAUWebshop conversion maintained
PickingOrder lines / hourNo impact on delivery delay

The structure is still in motion — some cells are still establishing their North Star because the measurement pipeline isn't there yet. We'd rather tell you that than pretend.

How we structure Rohlik Tech