Human-in-the-loop Pilot-ready MVP Not a placement decision

Extract, review, then shortlist — the caseworker decides

Foster Match helps UK fostering services turn carer and child narratives into structured profiles, require human approval, then rank approved carers for a referral. Outputs are decision support. Statutory placement stays with the caseworker.

Carer onboarding Caseworker portal Research bibliography →

What we are — and what we are not

  • ✓ Decision support — ranked shortlist with hard excludes, counter-arguments, and source quotes on extracted fields. Not an automated placement.
  • ✓ Literature-informed dimensions — scoring uses static weights we chose; they are not fitted to UK placement outcomes.
  • ✓ Sits beside CMS tools — CHARMS, Liquidlogic, Mosaic, and Link Maker remain systems of record. We do not replace them, and live two-way CMS sync is not shipped.
  • ✗ Not a dating-app filter — hard gates for law and safety only; soft scores for fit. No calibrated stability prediction (PSI).

Who it is for

Foster Match sits between a recruitment or referral narrative and a shortlist the placement team can review. It is not a vacancy marketplace and not a case-management system.

♥

Prospective foster carers

Write in your own words. The service structures a draft profile. An assessing social worker must review it before it can be used in matching.

Open carer portal →
◈

Placement and family-finding teams

Ingest referrals, review extracted fields, run a ranked shortlist of approved carers, see hard excludes and a devil’s-advocate panel, and record a decision if you select someone.

Open caseworker portal →

How it works

Profiles used in matching must be approved by a caseworker. The shortlist is advisory.

01

Narrative capture

Carer interview text or a caseworker document. Raw text is stored for review.

02

Structured extraction

An LLM maps narrative to a JSON schema. Uncertain fields are flagged. Critical values need a source quote before approval.

03

Caseworker review

An assessing SW approves or corrects fields. Unapproved carer profiles are not matching candidates.

04

Compatibility shortlist

Hard gates, then a static weighted score. Semantic ranking is off. Risk flags are heuristics, not a predicted outcome.

What we claim — and what we do not

A public evidence pack and bibliography sit with the API. Match responses include not_a_placement_decision: true. We do not claim UK outcome validation, SHAP-style score breakdowns in the UI, or that OpenAI extraction is a fact without review.

Research bibliography DPIA & data map
Not in this product today
  • Calibrated PSI — no placement-stability probability from UK outcomes
  • Live CMS poll — DBS / Terms gates only when those fields are already set
  • Semantic layer — embeddings may be stored; they do not rank the shortlist
  • National vacancy search — that is Link Maker territory, not ours
Live: ingest, review, shortlist MVP

Pilot-ready decision support — not a validated matching engine

What you can run today: narrative onboarding, extraction with confidence notes and evidence quotes, a review queue, a ranked shortlist with hard excludes and counter-arguments, and a decision record. Face validity with practitioners is how we will judge usefulness — not outcome statistics we do not have.