S solven_labs
AI automation agency · Manchester & UK

Agents that hold up in production.

We dissolve your problems. Cut through complexity.

Solven Labs is a UK AI automation agency that builds the other thing: custom agentic systems with proper tool use, real integrations into the stack you already run, and evaluation pipelines so they don't quietly fail at 3am. Not n8n flows dressed up as transformation.

Stack
Claude · MCP · custom tools
Practice
Evals before shipping
Founded by
An engineer, not a deck
doc-review-agent · tracing
§ The market, honestly

The gap between an automation and a system that runs your business.

Plenty of agencies will sell you the first one and call it AI. The second one needs engineering - the kind that catches its own regressions and survives a real Tuesday.

What agencies sell
-n8n or Make canvas with twelve nodes
-A Zapier chain that calls an LLM once
-Prompts pasted into a Google Doc as 'the system'
-No tests. No evals. Hope.
-Breaks when the input shape changes
-Hands you a Notion page and disappears
What Solven Labs builds
+Agentic systems with planned tool use and retries
+Real integrations: your CRM, case system, S3, internal APIs
+Prompts versioned in code, reviewed like any other change
+Eval suites against your data, run on every change
+Observability - traces, costs, failure modes - wired in
+Deployed into your stack. Owned, not held hostage.
§ What we actually build

Four shapes of project we tend to take on.

The brief usually starts as "we want AI to do X." The work is figuring out the smallest end-to-end slice of X that earns its keep, building that, and then expanding only where the evals say it's safe to.

Doc review

An agent that reads 800-page case bundles and finds the contradictions.

Claude with custom retrieval over a client's case management system. Tool use for PDF extraction, citation lookup, and report drafting. Evaluation harness against ground-truth examples reviewed by a senior solicitor.

ClaudeCustom retrievalPDF toolsEval harness
Internal ops

A triage agent sitting in front of a support inbox.

Reads incoming mail, classifies against a live schema of issue types, pulls account context from the team's Postgres, and either drafts a reply or routes to a human with the right notes attached. Wrong-classification rate tracked weekly.

MCP serverPostgresInternal APISlack handoff
Sales / RevOps

Outbound research agent that writes the first paragraph an SDR doesn't have to.

Pulls from the team's CRM and a few licensed data sources, builds a structured profile, then drafts an opener tied to something specific and recent. Prompt versions reviewed in pull requests; outputs scored against a rubric the head of sales actually agrees with.

CRM integrationStructured outputsPrompt versioningRubric evals
Compliance

A policy-aware reviewer that pre-checks documents before they go to legal.

Indexes a client's policy library, scores incoming material against it, and writes a short memo explaining where it would fail review and which clause to cite. Built so legal trusts it enough to use, not enough to be replaced by it.

RetrievalStructured gradingAudit logHuman-in-the-loop

Project shapes shown without client names. Specifics under NDA; happy to talk through them on a call.

§ Approach

How a project actually goes.

Five steps, in order. Engagements run six to twelve weeks for the first slice; longer if the integration surface is wide.

Duration
6-12 weeks
Delivery
In your stack
01

Scope the actual workflow

Sit with the people doing the job. Watch the boring parts. The agent's job description comes from that, not from a wishlist.

02

Build a thin slice end-to-end

One narrow path, real inputs, real tool calls, real outputs. No mocks past day three. The point is to hit reality early.

03

Evaluate against your data

Build an eval set from real cases the team has already handled. Score every prompt and model change against it. Regressions get caught before deploy.

04

Integrate into the stack you run

Whatever you already use - Postgres, S3, HubSpot, your case system, your auth - that's where the agent lives. No new dashboard for the team to learn.

05

Hand over with observability

Traces, cost, failure modes, eval scores: all wired to something you can read at 9am on Monday. You own the system. We're available when you need us, not when you don't.

§ Who runs it

Marcin. Software engineer, builder of regulated software.

Solven Labs is run by Marcin, a software engineer and co-founder of ALLDOQa UK SaaS platform in the medico-legal space, used by HM Courts & Tribunals, instructing solicitors, and around 180 medical experts. ALLDOQ is double-encrypted, OWASP-aligned, and held to the standards a court bundle requires. The same engineering discipline that makes a platform safe enough for a courtroom is what's wanted in an agent that's allowed to touch production data. That's the work here.

Based
Manchester, UK
Background
Regulated SaaS
Working hours
UK · async-friendly
// latest thinking
§ Questions

Questions people ask
before they get in touch.

What does Solven Labs do?

Solven Labs is a UK AI automation agency that builds custom AI agents and automation systems for businesses. The work centres on real integrations into the stack you already run, with evaluation pipelines so the system keeps working in production rather than failing quietly.

What makes Solven Labs different from a typical AI automation agency?

Most agencies sell a workflow canvas with a few nodes and call it AI. Solven Labs builds engineered systems with versioned prompts, eval suites run against your own data on every change, and observability wired in, then deploys them into your stack so you own them. The approach is engineering-led rather than demo-led.

How much does an AI automation project cost?

A pilot that proves one narrow process tends to run between 5,000 and 15,000 pounds, and a production build of a single full process sits between 15,000 and 45,000 pounds for most UK SMEs. The figure depends on how many systems it touches, how messy the inputs are, and the cost of a wrong answer. Full breakdown in the cost guide.

How long does an AI automation project take?

The first slice usually runs six to twelve weeks, and longer when the integration surface is wide. The aim is the smallest end-to-end slice that earns its keep, expanded only where the evals say it is safe to.

Where is Solven Labs based and who runs it?

Solven Labs is based in Manchester and works with businesses across the UK. It is run by Marcin Walczak, a software engineer and co-founder of ALLDOQ, a UK medico-legal SaaS platform held to the standards a court bundle requires.

§ Contact

Start a conversation.

Most useful when you include the rough shape of the workflow, what data the agent would need to touch, and any constraint - regulatory, latency, or otherwise. Replies within two working days.

Please enter your name.
Please enter a valid email address.
Please enter a message.
// message sent - we'll reply within two working days.