Why your RAG system fails on the twentieth question
Retrieval quality degrades quietly as a corpus grows. A practical guide to chunking strategy, hybrid scoring and the evaluation set that catches the regression before your users do.
Read the noteWasabiLabs is an applied AI studio. We design, ship and operate production systems — autonomous agents, forecasting engines, computer vision and document intelligence — for companies whose margins depend on getting the details right.
# production agent — claims triage, EU region pipeline: claims-triage-v7 runtime: orchestrator: wasabi-runtime region: eu-west-1 residency: eu-only models: - role: extraction latency_p95: 410ms - role: reasoning latency_p95: 1.2s - role: verification latency_p95: 380ms guardrails: groundedness: 0.96 # min. score pii_redaction: strict human_review: below_confidence(0.88) observability: traces: full evals: nightly + shadow
Trusted by operators in regulated and high-throughput industries
Most AI projects die between the demo and the first real user. We organise around the disciplines that close that gap — modelling, engineering, evaluation and operations — so the thing you approve in a workshop is the thing that runs on Monday.
Multi-step agents that read, decide and act inside your stack — with tool access, memory, escalation paths and a human in the loop where it matters.
Demand, price, risk, capacity and maintenance models built on your own history — paired with solvers that turn a prediction into a decision someone can act on.
Inspection, counting, safety and quality models that run at the edge, on the line, in the yard or on a drone — offline-tolerant and calibrated to your defect taxonomy.
Contracts, claims, invoices, dossiers and scans turned into structured, auditable data — with citations back to the source page and confidence on every field.
The unglamorous half of AI: pipelines, lineage, chunking strategy, embeddings, hybrid search and permissions that mirror the ones you already enforce.
Evaluation harnesses, red-teaming, drift alarms, cost ceilings and audit trails. If it cannot be measured on a dashboard, we do not call it done.
A fraud model and a defect model share almost no vocabulary. We keep standing teams per sector so the conversation starts at your workflow, your regulator and your failure modes — not at a generic architecture diagram.
Models that have to explain themselves. We build under the assumption that every decision will eventually be read out loud to a regulator, an auditor or an unhappy customer.
Graph-aware scoring with narrative explanations attached to each alert, so analysts spend their time deciding rather than reconstructing.
RiskIdentity documents, proof of address and corporate structures parsed, cross-checked and routed — with straight-through processing above a tuned confidence floor.
OperationsAlternative-data features and reason codes that slot into existing scorecards without breaking model governance.
LendingPortfolio context, product rules and suitability constraints in one assistant that drafts client-ready notes with citations.
Front officeContinuous ingestion of circulars and consultations, mapped to the internal controls each one touches.
ComplianceIntent, severity and redress classification across channels, with automatic detection of systemic issues before they become a thematic review.
ConductClinical-grade caution with startup delivery speed. Everything is built assuming PHI never leaves the perimeter and every output is traceable to a source.
Ambient note drafting that respects local templates, coding rules and the clinician's own editing habits.
Care deliveryPolicy criteria matched against the chart, with the supporting evidence assembled into a submittable packet.
Revenue cycleWorklist prioritisation and second-read flags for radiology and pathology, deployed as decision support with clear uncertainty bands.
DiagnosticsAdverse event detection from case narratives, literature and unstructured intake, coded to MedDRA with human sign-off.
SafetyProtocol criteria turned into executable cohort queries across EHR and registry data, with site-level yield estimates.
R&DMultilingual assistants for scheduling, preparation and follow-up that escalate cleanly to a human at defined clinical boundaries.
ExperienceMargin lives in a thousand small decisions per day. We automate the ones that are mechanical and sharpen the ones that are judgement calls.
SKU × store × week forecasts with promotion, weather and cannibalisation effects, reconciled across the planning hierarchy.
PlanningElasticity-aware price recommendations bounded by brand, competitive and legal guardrails.
CommercialAttribute extraction, deduplication and copy generation across millions of listings in every market language you sell in.
ContentAssistants that know stock, delivery promises and return policy — and hand over with full context when they should.
ServiceBehavioural scoring that separates genuine dissatisfaction from organised abuse, without punishing good customers.
RiskShelf availability, planogram compliance and queue analytics from existing camera infrastructure.
StoresShop-floor reality beats laboratory accuracy. We deploy to the hardware you already have, tolerate bad lighting and patchy networks, and measure success in scrap and downtime.
Defect detection tuned to your taxonomy, retrained continuously from operator corrections on the line.
QualityVibration, thermal and telemetry fusion that gives maintenance a ranked worklist, not another alarm to ignore.
ReliabilityClosed-loop recommendations for setpoints that hold yield while cutting energy and cycle time.
ProcessDecades of manuals, work orders and tribal fixes made searchable in the language technicians actually use.
KnowledgeEarly signals on supplier distress, lead-time drift and single-source exposure across the bill of materials.
Supply chainPPE compliance and exclusion-zone monitoring designed with works councils, privacy-preserving by construction.
EHSNetworks where a five-minute decision delay compounds across a continent. Our systems run at the edge of the network and degrade gracefully when the link drops.
Fleet, driver-hours and customer-window constraints solved continuously, not once per night.
PlanningLive traffic, dwell-time and handover models that replace static transit tables with honest probabilities.
VisibilityBills of lading, customs forms and proof of delivery extracted and reconciled against the transport order.
Back officeTrailer identification, occupancy and damage capture without new gate hardware.
TerminalDelays, damages and misroutes detected, explained and actioned with the carrier before the customer notices.
OperationsScenario modelling for hub placement, mode shift and capacity commitments under demand uncertainty.
StrategyPhysics-constrained models for assets that cannot be A/B tested. We work alongside your engineers, and we never ship a recommendation that violates the grid code.
Weather-ensemble driven forecasts at asset and portfolio level with calibrated uncertainty for trading desks.
MarketsSensor-level deviation detection that distinguishes instrumentation faults from genuine network events.
NetworksDrone and line-crew imagery triaged for corrosion, vegetation encroachment and component wear.
FieldBattery and demand-response scheduling against price, constraint and degradation objectives.
OptimisationTariff, consumption and efficiency guidance grounded in the customer's own meter data.
RetailEnvironmental and regulatory submissions assembled from project data with full traceability.
DevelopmentWhere a hallucinated citation is a career event. Every answer is grounded, quoted and linked, or the system says it does not know.
Clause extraction scored against your fallback positions, with redlines drafted in your own house style.
TransactionalData-room triage that surfaces change-of-control, indemnity and unusual terms before the deadline, not after.
M&APrecedent and memo retrieval that respects ethical walls and matter-level permissions.
KnowledgeObligation extraction and impact mapping across jurisdictions your clients actually operate in.
AdvisorySampling, tie-out and workpaper drafting with provenance to the underlying ledger entry.
AssuranceFirm credentials, prior matters and pricing precedent composed into a first draft in minutes.
GrowthAccountability is the product. We build for procurement scrutiny, freedom-of-information requests and the day a journalist asks how the system decided.
Multilingual, accessible guidance on entitlements and procedures, grounded strictly in published policy.
ServiceApplication routing and completeness checks that shorten queues without removing human judgement from outcomes.
AdministrationTender comparison, spend concentration and anomaly detection across framework agreements.
FinanceRetrieval, redaction and response drafting with statutory deadlines tracked end to end.
TransparencyCurriculum-aligned tutoring and formative feedback with strict age-appropriate safeguards.
EducationRoad, rail and utility condition surveys from vehicle-mounted and satellite imagery.
WorksEvery engagement stands on the same foundation: a modular runtime we maintain, deploy inside your cloud or ours, and hand over with the source. No black boxes, no lock-in, no surprise per-seat bill in year two.
Connectors, change-data-capture and normalisation across the systems of record you already run — with lineage preserved end to end.
Hybrid vector and lexical search, structured joins, and permission inheritance so the assistant can never surface what the user could not open themselves.
Frontier, open-weight and in-house models routed per task on quality, latency and cost — swappable without touching business logic.
Typed tools, deterministic state machines around probabilistic steps, retries, compensating actions and explicit human checkpoints.
Offline evals, online shadow scoring, groundedness checks, cost ceilings, drift alarms and a complete trace for every decision made.
Your VPC, our managed estate, on-premise or the factory floor. Infrastructure as code, reproducible builds, and data that stays where your counsel says it must.
We do not run open-ended discovery. Every engagement starts with a decision we are trying to change and a number we expect to move, and ends with that number on a dashboard your CFO can open.
Two weeks. We map the workflow, quantify the decision, and agree what "good" looks like before any model is trained.
Four to six weeks. A working slice on your real data, evaluated against a golden set your experts built with us.
Six to twelve weeks. Hardening, integration, security review, change management and a staged rollout behind flags.
Ongoing. We run it, watch it, retrain it and keep proving the value — or we hand the keys to your team and stay on call.
When we solve the same problem for the fourth client, we turn it into a product. Each one ships with its own deployment, its own pricing and its own landing page — and each one is still maintained by the team that built the original.
Document intelligence
Turns contracts, claims and dossiers into structured, cited data. Field-level confidence, human review routing and straight-through processing out of the box.
Forecasting
Hierarchical demand and capacity forecasting with probabilistic outputs, promotion effects and automatic reconciliation across planning levels.
Agent platform
The orchestration layer we use internally: typed tools, deterministic workflows around model calls, replayable traces and built-in human checkpoints.
Industrial vision
Edge-deployed inspection for production lines and yards. Ships on commodity hardware, learns from operator corrections, works when the network does not.
AI assurance
Evaluation and monitoring for AI systems in production: golden sets, regression gates, groundedness scoring, drift alarms and EU AI Act evidence packs.
Knowledge & search
Enterprise retrieval that inherits your existing permissions. Hybrid search across documents, tickets and databases, with answers that always cite the source.
Every figure below was measured by the client's own finance or operations team, against a baseline we agreed before the work started.
Insurance · Europe
Eleven thousand daily FNOL submissions across five languages, previously read by hand. We built an extraction and severity pipeline with automatic fraud flags and a confidence-gated fast lane.
Manufacturing · DACH
A pilot cell proved out in six weeks, then rolled to every plant on existing industrial cameras and on-premise GPUs. Operators label corrections inline; the model retrains nightly.
Retail · Iberia & LATAM
Replaced a spreadsheet-driven planning process with hierarchical forecasts, promotion modelling and automatic order proposals that planners approve rather than assemble.
Healthcare · Nordics
Deployed entirely inside the hospital's own data centre. Clinician-authored templates, local language models, and a hard rule that nothing is filed without a human signature.
Logistics · Global
An agent that watches milestone events, detects at-risk shipments, drafts the carrier chase and books the recovery option — escalating to a human when cost exceeds a threshold.
Banking · UK
A single evaluation and documentation pipeline covering 40+ models, cutting the regulatory submission cycle from quarters to weeks with evidence generated automatically.
We had two failed AI pilots behind us. WasabiLabs spent the first week telling us which of our ideas would not work. That is why the third one shipped.
They handed over the repository, the runbooks and the on-call rota. Six months later our own team is shipping changes without them. That was always the deal.
Our regulator asked for the decision trail on a Friday afternoon. We exported it in twenty minutes. That single moment paid for the engagement.
We work in sectors where a wrong answer has consequences. Safety, privacy and auditability are engineering requirements in our backlog from sprint one — priced in, not bolted on.
Your data stays in the jurisdiction you nominate. Air-gapped and on-premise deployments are a standard option, not an enterprise upsell.
Every output stores its inputs, retrieved evidence, model version and prompt. Replay any decision from any date, for as long as your retention policy requires.
We define the confidence thresholds and failure modes that must reach a person, and we design the review interface so that person can act in seconds.
Prompt injection, data exfiltration, jailbreak and bias probes run on every release — with the results written into the release notes your auditors read.
WasabiLabs was founded in 2019 by engineers and researchers who were tired of AI projects that ended in a slide deck. Forty-six people across Barcelona, Amsterdam and Tallinn. No account managers between you and the people writing the code.
Co-founder & CEO
Fifteen years in applied ML, previously led decision science at a European payments network.
Co-founder & CTO
Distributed systems and MLOps. Built the inference platform behind a 40M-user marketplace.
Head of AI Research
PhD in probabilistic forecasting. Publishes on calibration and uncertainty in production models.
Head of Delivery
Runs the engagement model. Previously delivered core banking migrations across three continents.
Monitoring, permissions, retries and rollbacks are what separate a demo from a system. We build those first and treat them as features, not overhead.
We never train shared models on your data. Contracts assign the work product to you, source included, from the first commit.
If we think a use case will not repay its cost, we say so in the first workshop. Roughly a third of our proposals recommend a smaller scope than the one we were asked to quote.
Written by the engineers on the projects, not by a content agency.
Retrieval quality degrades quietly as a corpus grows. A practical guide to chunking strategy, hybrid scoring and the evaluation set that catches the regression before your users do.
Read the noteWhat high-risk classification means in engineering terms, and how to generate most of the required documentation from artefacts your pipeline already produces.
Read the noteToken cost is rarely the problem. Routing, caching, retries and the cost of a wrong answer are — with the arithmetic we use to size an engagement.
Read the noteTwo weeks for a framed problem with a measurable baseline, four to six more for a working slice on your own data. We do not run six-month discovery phases; if a use case needs one, it is not ready to be an AI project yet.
Usually not. Most of our engagements begin with data that lives in an ERP, a shared drive and three spreadsheets. We build the minimum pipeline the use case needs and leave it behind as reusable infrastructure rather than asking you to finish a two-year platform programme first.
Where you tell us. We deploy into your cloud account, our managed EU estate, your own data centre or a fully air-gapped environment. We never train shared models on client data, and every deployment ships with a data flow diagram your DPO can sign.
Whichever fits the task. A frontier model for open-ended reasoning, a fine-tuned small model for high-volume classification, gradient boosting for tabular problems. Model choice is a routing decision inside the runtime, so swapping one costs a config change rather than a rewrite.
We design for it. Confidence thresholds route uncertain cases to a person, guardrails block actions outside a defined envelope, and every decision is replayable. We agree the acceptable error profile with you before launch and monitor against it continuously.
Fixed price per phase for Frame and Prove, so you can stop cleanly after either. Ship is time and materials against an agreed scope, and Operate is a monthly retainer tied to SLOs. For some products we also offer outcome-linked pricing once a baseline is established.
You do. Source code, models, training data and documentation are assigned to you. Our platform components are licensed perpetually and irrevocably for the deployed use case, with the source escrowed. There is no scenario in which turning off a WasabiLabs subscription turns off your system.
That is our preferred mode. We embed two to five people alongside your engineers, pair on the work, and plan the handover from day one. About 60% of our engagements end with the client's own team operating the system independently.
Send us one workflow that costs you money or time. Within five working days you will get a written assessment: whether AI is the right tool, what it would take, what it would cost, and what could go wrong.