Sekhmet

…

seed / resolved person company

Click any name to find them in the graph.

Aleph (OCCRP)Planned

NorthData (Europe)Phase 2

OpenCorporates (paid)Phase 2

IngestPull officers, PSCs, addresses for the seed and recursively expand through co-directors.

In demo

ResolveMerge multiple registry records for the same individual (DOB + name-token + address overlap).

Hand-curated

ExpandBipartite graph traversal to depth K. High-degree nodes (corporate secretaries, large partnerships) are capped to prevent graph explosion.

In demo

EnrichAttach per-node signals from external sources (sanctions, leaks, electoral roll, public records, KYC). Each source is a pluggable adapter.

In demo

Kinship detectionFamily and household relations: inferred from corroborating signals (joint directorships, reciprocal electoral-roll co-occupancy, surname overlap at the same outward postcode). Paid KYC and full-address household-occupant feeds extend the coverage in Phase 2.

Basic in demo

Cross-referenceIntersect the network with the client watchlist. Surface flagged conflicts.

Phase 1

OutputCSV (preserving the analyst's existing workflow) and interactive graph view, with intersection-flag column highlighting hits.

In demo

ExtensibleThe pipeline is modular: additional stages and source adapters plug in alongside the above without changes to the core.

Pluggable

What this prototype is, and what it isn't

The prototype validates the engine shape against real public data on two targets. Data fetches are ad hoc rather than batch-pipelined, and the matchers in some stages are rule-based defaults rather than learned models. Phase 1 turns these into the production engine. Roughly in order of effort:

Persistent fetch layer. Checkpointing per seed and per source, on-disk caching, rate-limit-aware retry, schema validation, structured logging.

Phase 1

Entity resolution at scale. Replace the rule-based heuristic with a calibrated probabilistic matcher emitting per-decision confidence.

Hand-curated

Batch runner. Handle hundreds to thousands of seeds reliably with checkpointed state.

Phase 1

Stop-rule tuning. Replace inspection-set thresholds with values learned from corpus-wide degree distributions; expose as per-deployment tunables.

Defaults set

Watchlist ingestion + intersection. Handle client format, fuzzy-match thresholds, per-conflict provenance.

Phase 1

Additional source adapters. Aleph and other planned sources, plus ongoing maintenance as endpoints change.

Some done

Operational tooling. Alerting on source breakage, run-health dashboards, cost monitoring.

Phase 1

The visualisation, source-adapter pattern, and methodology in the prototype carry directly into the Phase 1 build.