Original development · Local reference implementation · October 2026
Connector Forge
A security-data workbench that reconciles observations, preserves provenance, and puts connector changes through evaluation and human review.
The problem
Security feeds can repeat records, arrive out of order, and disagree. A useful inventory needs consistent merge behavior and a way to explain where each field came from.
My contribution
I built the Python/SQLite ingestion core, service integration, evidence curation, local dashboard, and reviewed mapping lifecycle. Six local services separate source generation, reconciliation, curation, advisory AI, mapping evaluation, and cycle coordination.
Interactive demonstration · invented records
Arrival order should not rewrite history.
Two observations disagree about the same asset. The newer non-empty field wins; both source references remain available. This browser demonstration illustrates the documented merge rule, not a running backend.
hostname: lab-oldsource: record-ahostname: lab-currentsource: record-bRollback restores a previous adapter mapping as a new generation; it does not erase observations or undo their ingestion. This example starts with an approved candidate mapping.
Engineering decisions
Each mutable field has its own precedence clock. A newer non-empty assertion replaces an older one; source arrival order is not the decision rule. Composite finding identities keep distinct package versions separate.
Service-owned databases avoid cross-service writes. Mapping candidates are declarative data, evaluated against golden cases and freshly checked before human-approved activation. The optional AI gate is advisory and cannot change records, policies, or approved mappings.
Evidence & validation
Retained local verification notes cover contract cases, rejected input with no partial mutation, duplicate-safe replay, reversed arrival order, mapping evaluation, review, and rollback. Requirements map to behavior checks; that mapping is distinct from line or branch coverage.
The website demonstration exercises a deliberately small illustration of precedence, replay, provenance, and mapping history. It does not execute the private connector or replace its verification suite.
What the work demonstrates
The implementation makes reconciliation and controlled change inspectable: follow a projected record to its source, reproduce the processing decision, and review a mapping before changing active behavior.
Scope & limitations
The source observations are synthetic. This is a single-tenant local reference implementation, not validated vendor detection or production infrastructure.
Evidence curation and mapping evolution do not train model weights or generate and execute arbitrary code. Production identity, transport security, independent audit storage, and operational hardening remain separate work.
What you can inspect
The diagram, interactive example, and implementation account are public. The source and retained local check records were reviewed for this portfolio; private code, operational records, and notification destinations are not published.
Get in touch
Have a challenge
worth solving?
Start a conversation about engineering, research, or a problem that needs a closer look.