Articles
ComplianceGovernanceDeadlinesProcurement rules tracker: the dates that changed what a bid has to do
A dated tracker of three procurement rule changes: the UK Procurement Act 2023 going live, eForms becoming mandatory for EU notices, and the European Commission proposal to replace the public procurement directives.
AutomationWorkflowVerificationAgentic bid workflows: where an agent helps and where a person signs
What it would take for an agent that drafts, checks and assembles bid material to be verifiable, where autonomy helps, and why the human must hold the pen on a legal commitment.
ModellingClassificationDataProcurement as a graph: buyers, suppliers, awards and the classifications between them
Why buyers, suppliers, awards and classification codes form a graph, why joining CPV to DK021 is harder than it looks, and what is still open in the work we are doing on it.
MatchingCapabilityScoringKeyword alerts are noise; a capability profile is a filter
Why keyword alerts drown bid teams, and what it means to score notices against what a firm can actually deliver, with the reasons shown.
AuditVerificationSecurityA bid record is evidence, so someone who does not trust us must be able to check it
Why a bid record has to be checkable by an outsider, who that affects, and which parts of the Klarum record can be checked without trusting us today.
DocumentsExtractionDataTender documents are not text, and extraction has to know it
Tender packs arrive as PDFs, spreadsheets and scans whose meaning lives in their structure. What structure-preserving extraction is, who it affects, and where it still fails.
ResearchSecurityAuditEncryption at rest and a partly verifiable audit trail
What Klarum encrypts, who holds the key, and which part of the audit trail an auditor can check without trusting us. Summarises our security whitepaper for vendor evaluation, with the document attached.
ScoringMatchingMethodKlarum match scores: ten calibrated signals, with missing data redistributed
The ten signals and weights behind a Klarum match score, why an uncomputable signal is redistributed rather than scored as a poor fit, and what the score does not measure. No evaluation of the ordering has been run.
SourcesDataCoverageInside the Klarum tender corpus: sources, normalisation and coverage
The public sources behind the Klarum corpus, what normalisation destroys and preserves, how two incompatible sector taxonomies are bridged, and the limits of the coverage figures.
PlatformWorkflowRecordsOne record, two directions
Discovered tenders and received applications enter from opposite directions and converge on one bid workspace. The case for a single pipeline, and what the convergence costs on the inbound side.
ResearchExtractionCapabilityCapability extraction: every claim tied to its source passage
How we extract what a firm can prove it has done from its own documents, why every capability carries the passage that evidences it, and where a controlled vocabulary helps and hurts.
ResearchReferencesRankingReference selection for bids: ranking past projects with no ground truth
Reference selection for a bid is a ranking problem with no ground truth. Candidates are ranked by similarity, discounted for age and reranked on four named factors, and the objective optimised is a proxy for the one that wins bids.
What we work on
An article says how something works. This says whether you can use it today.
- Semantic document extractionIn productionTender documents arrive as PDFs, spreadsheets, and scans whose structure plain text parsing destroys. We work on extraction that preserves meaning: tables, cell colors, requirement hierarchies, CV structure.
- Capability-aware matchingIn productionKeyword alerts drown teams in noise. We model what a firm can deliver, from its documents and win history, and score notices against that profile with reasons attached.
- Verifiable records and confidentialityIn productionA bid record is evidence, so it has to be checkable by someone who does not trust us. We work on tamper-evident audit chains, signed roots, and receipts an auditor can verify offline with no involvement from Klarum, alongside encryption that keeps customer documents readable only by the people who should read them.
- Procurement knowledge graphsActive workBuyers, suppliers, awards, and classifications form a graph that predicts who wins what. We build and query that graph across sources and classification systems such as CPV and DK021.
- Agentic bid workflowsActive workAn agent that drafts, checks, and assembles bid material has to be verifiable. We study where autonomy helps and where the human must hold the pen.