B2B logistics marketplace platform
We were engaged to build a marketplace matching businesses needing warehousing and transport with providers that have capacity.
The market and the problem
We built it for a UK logistics client we can't name. Once a match is made, the platform runs the relationship end to end: stock, despatch, transport, invoicing, settlement. Every change has been kept as a permanent record since day one, and five AI features run on it in production.
Contract logistics is a long-tail market. Most warehousing and haulage capacity in the UK sits with small regional operators, which makes it expensive to find and slow to buy. Sourcing a fulfilment partner routinely takes months of calls, site visits and negotiation, and usually ends in a multi-year commitment that doesn't flex with seasonal volume.
The supply side has the mirror problem. Operators sit on idle racking and empty return legs, and the cost of winning a new contract through traditional sales is brutal for a business running on thin margins.
Marketplaces are the obvious answer, and mostly they stop at the introduction. The interesting decision here was to keep going: once shipper and provider contract, the same platform runs the operation, from goods-in to proof of delivery to the invoice. That is what turns a directory into infrastructure, and it is why the build is as large as it is.
What we built
We delivered one platform with two sides and a workflow that follows the relationship from first search to settled invoice.
Search that starts anonymous
Public geo-radius search across more than twenty facets, from temperature control to certifications to racking dimensions. An AI concierge qualifies what the visitor needs, shows anonymised matches, and turns interest into a registered, scored lead.
Contract negotiation and e-signature
Fourteen service rates negotiated individually, each carrying the provider's list price, both sides' proposals and the agreed figure, plus volume-discount tiers. E-signature is wired into the flow, and the contract state machine won't let a deal skip steps.
Fulfilment operations
A fulfilment workflow from despatch note to proof of delivery, enforced server-side, with stock moving through allocated and in-transit counters at every step. Both parties see the same process, trimmed to what each is entitled to see.
Transport quoting inside a B2B logistics marketplace
Quotes priced from multiple rate-card sources with postcode-zone matching, capacity and dimension gating, and honest labelling: a quote is either firm or indicative, and each carries a cost, capacity and directness score.
Marketplace settlement and invoicing
Billing cycles per contract, proforma to finalised invoices with the full margin stack itemised, and payment initiated over open banking. Money flows are part of the platform, which is what makes the commission model enforceable.
Vetting before go-live
A listing only appears in public search once vetting passes, and the vetting is done by an AI classifier with domain sense. Suggestions go back to the provider; nonsense never reaches a buyer.
The engineering
This is the largest platform we have built: a .NET backend, two front ends and a fleet of AI services. Two decisions shaped everything else.
The platform in outline. The marketplace app and the admin console use the backend. Every write appends a permanent, hashed event to the event store, and Elasticsearch runs the search. Over the programme it ran on Kubernetes, hosted on AWS, Google Cloud and Azure, as the line under the drawing shows. Highlighted: the event store, which keeps every change as a permanent record.
What happens when two companies disagree about a shipment?
The platform can show exactly what happened, in order, and that the record has not been altered since. Every write appends a permanent event whose contents are hashed, and the hash is checked again whenever the history is replayed. Versioned migrations let the data model change without ever breaking that history. A clashing edit is rejected by the hash check rather than blocked with a lock, which sounds brave and has proven simple to reason about.
Shared records and private rate cards
Both parties to a shipment work from the same record, and neither sees what belongs only to the other. Standard SaaS isolation gives each customer its own separate world, which a marketplace cannot do: shipper and provider must both see the shared record, each trimmed to their side. Quoting goes further, pricing a shipper's request against many providers' private rate cards through a deliberately hidden reader that cannot leak any of them.
The card "Shared records and private rate cards", drawn. Both parties to a shipment work from the same record, and neither sees what belongs only to the other: the shipper and the provider must both see the shared record, each trimmed to their side. Highlighted: the shared record.
Search and matching
Shippers find space by distance and by what a warehouse can handle, even from a mistyped postcode, and search stays up while it is rebuilt. Elasticsearch runs the radius search, the filters and the postcode matching; a rebuilt index is prepared under a new name and swapped in, so a rebuild never takes search down.
Self-hosted services
Fewer outside companies handle the client's commercial data, and no per-seat bills grow with success. Notifications, e-signature, PDF rendering and secrets management all run inside the platform's own cluster, and error tracking is an in-house system that joins front-end and back-end errors into one view.
Operated as infrastructure
Stock, shipments and money all move through this platform, so it is tested and operated like infrastructure. Over the programme it ran on Kubernetes, hosted on AWS, Google Cloud and Azure. Automated tests cover the backend, and an admin console gives the operations team tools to browse and replay the event history.
Where AI fits
This platform runs more AI in production than anything else we've built. Logistics is a trust trade with a paperwork problem, and paperwork is where current models are at their most useful.
The public assistant's model step, inside its grounding guards. One guard strips any result the model presents without a matching tool call, so the assistant cannot invent a warehouse, and the chat still answers when a guard fires. Highlighted: the model step.
- 1 to 6: the grounding guards that sit behind the assistant.
The AI concierge
The public assistant routes each visitor to a specialist agent for warehouse search, transport quotes or onboarding. Six grounding guards sit behind it, and one of them strips any result the model presents without a matching tool call, so it cannot invent a warehouse. Errors never surface as failures; the chat always answers. Leads arrive in the CRM scored for intent.
Ongoing model evaluation
Three generations of the assistant have shipped, running across several providers with weighted A/B testing, an LLM-as-judge evaluation rig with named scenarios, and per-model cost accounting. Which model answers a given turn is decided by measurement.
AI listing vetting
The vetting classifier checks each new listing against domain rules (coordinates that exist, plausible dimensions, credible pricing), and its verdict decides whether the listing goes live at all.
Ready for customers' AI agents
An MCP server exposes the whole operational API to AI agents, each tool classified read, create, mutate or destructive. When a customer's agent needs to check stock or raise a despatch note, it can do so through those tools.
Phases, team and price
Delivery record
- Status
- Live in the UK market
- Programme
- More than three years, in phases
- Team
- Engineers across phases
- Hosting
- AWS, Google Cloud and Azure
Case study published , updated .
A programme measured in years, structured so each phase paid for the next. The first transactable phase has its own row in our product development price bands.
Four phases, in the order listed below. Highlighted: phase 1, marketplace core, the first transactable loop.
Indicative price
£280k to £450k in year one
A first transactable marketplace of this shape typically runs £125,000 to £170,000. With an operations platform and continued iteration, first-year totals usually land between £280,000 and £450,000, and multi-year programmes of this depth scale well beyond that. Excludes infrastructure, at London rates as of mid-2026. Shape bands, not an invoice.
The bands in this study's price note, on one linear scale from zero and numbered in the key. Each is a band for the shape, not an invoice. Highlighted: band 2, the one this page leads with.
- 1. A first transactable marketplace: £125,000 to £170,000
- 2. First-year totals, with an operations platform and continued iteration: £280,000 to £450,000
How we build
What every build goes through
These commitments hold on every build we run:
- Nothing is built until the specification is written down and you have agreed it.
- A senior engineer reviews every change before it is merged.
- Automated tests run on every change, and a failing build does not ship.
The delivery record shows this project's own details. How we work sets out the rest of the method.
What the engagement achieved
The platform is live in the UK market with providers and buyers transacting through it, and the operations side, the part that was hardest to build, is the reason relationships stay on-platform once they start.
Much of logistics still runs on phone calls and PDFs. This platform vets its own listings and qualifies its own leads, and customers' AI agents can already work through its API.
- Marketplace and operations on one platform, from first search to settled invoice
- A hash-verified, replayable event store that keeps every change as a permanent record
- A fulfilment workflow enforced server-side, from despatch note to proof of delivery
- Five AI features in production, with an evaluation rig
- Automated tests across the backend, on a platform operated like infrastructure
Marketplace and operations on one platform, from first search to settled invoice: search, contract, fulfilment, then settlement and invoicing. Highlighted: fulfilment, the first step where, as the brief puts it, the platform runs the operation.
The decision we would defend
The assistant cannot invent a warehouse
One of six grounding guards strips any result the model presents without a matching tool call. A visitor may get a shorter answer, but never a warehouse the platform does not hold, and the chat still answers when a guard fires.
Planning something like this?
Indicative price
£280k to £450k in year one
A band for the shape, not an invoice.
The first step
Specification sprint
£4,500 or £8,500One to two weeks
The written specification, architecture, estimate and risks: one week for a single workflow, two for a platform. You can take it to any supplier, us included.
Asked about this project
Why is the client not named?
The engagement is under NDA, which is common for platform builds in competitive markets. The screens shown are indicative recreations in our own visual language, and identifying details, including scale figures the client publishes, have been generalised.
What does the platform do?
It matches businesses that need warehousing and transport with providers that have capacity, then runs the resulting relationship: contracting, stock, goods in and out, transport bookings, invoicing and payment. Marketplace plus operations platform in one system.
What AI runs in production?
An LLM concierge with grounding guards that qualifies visitors and scores leads, an AI classifier that vets listings before they go live, document extraction that turns rate-card PDFs into structured pricing, and an MCP server that lets AI agents operate the platform. The assistant runs across several models, compared under weighted A/B tests and an evaluation rig.
1 more question
What does a two-sided marketplace like this cost to build?
A first transactable version of this shape typically costs £125,000 to £170,000. A serious first year, marketplace plus operations, usually lands between £280,000 and £450,000 at London senior rates. The multi-year programme here went well beyond that, phase by phase.
Tell us what you're building or deciding.
Tell us about a product you want built, or a decision about AI you need to make, and we'll set up a 30-minute call. You'll get a straight answer on what it would take.
What happens after you get in touch
We reply within one working day
By email, to arrange a time for the call.
A free 30-minute call
With a senior engineer, about what you are building or deciding.
The fee in writing first
If a first step is worth taking, you get its exact fee in writing before any work starts.