HIRING / MULTI-AGENT VERIFICATION
HIRING PLATFORM
A multi-stage hiring system with provenance-aware evidence aggregation, deterministic five-pillar matching and Ed25519-signed JSON skill passports.
Persistent agent orchestration and verifiable skill credentials.
Fair Hiring Network decomposes candidate assessment into persisted, inspectable stages. ATS extraction and platform-specific services contribute claims and evidence; an evidence graph retains provenance and conflicts before skill-confidence, test-recommendation, bias-check and matching stages run. The resulting passport is serialised canonically and signed. Candidate, company and reviewer surfaces expose the workflow, while a separate Gemini Live path supports interview sessions and reporting.
Matching components
A returned breakdown accompanies the aggregate score.
Passport signature
Canonical JSON payload with an explicit verification endpoint.
Agent dispatch modes
Shared agent client across local and remote service boundaries.
Weighted matching model
C, F, X and E denote the normalised core, framework/tool, experience and evidence contributions; L is clamped learning velocity and λ is the role-specific learning velocity multiplier, producing a rounded aggregate score.
02 / THROUGH THE SYSTEM
INPUT → OUTPUT- 01
Collect
Applications connect résumé input with GitHub, coding-platform and professional-profile signals.
- 02
Analyse
Specialised stages extract claims, inspect evidence and flag suspicious or conflicting inputs.
- 03
Connect
An evidence graph preserves source provenance and feeds skill-confidence and test recommendations.
- 04
Match
A weighted, multi-component score connects the candidate’s evidence to a configured role.
- 05
Issue
A structured skill passport is signed with Ed25519 and can be checked through a verification endpoint.
Persisted orchestration across heterogeneous services
The orchestrator stores pipeline state and agent-run history with SQLAlchemy/PostgreSQL. AgentClient supports in-process ASGI dispatch and HTTP transport with exponential retries, while Flask-based services interface through WSGI integration. Stage checkpointing enables immediate resumption and auditing of completed verification tasks.
Provenance and adversarial-input screening
ATS processing extracts structured résumé claims and screens for hidden text and prompt-injection patterns. GitHub, LeetCode, Codeforces and professional profile services provide source-specific signals. Evidence-graph processing retains verified origins and conflict indicators, giving technical recruiters full traceability into matched skills and generated assessment rubrics.
03 / SELECTED EVIDENCE
04 FRAMES
Evidence-source acquisition
Candidate inputs feed separate verification stages. Mandatory profile/resume data and optional platform signals enter the normalisation and evidence-aggregation workflow with distinct provenance.

Signed skill-passport representation
The vertical credential displays assessed skills and aggregate metrics. The backend signs canonical JSON representations with Ed25519, allowing external parties to cryptographically verify passport integrity against the public key.

Role-specific matching output
The company view groups candidates under a configured role. MatchingAgent returns five weighted score components alongside matched/missing skills and an actionable decision band.

Gemini Live interview session
The interview interface coordinates live WebRTC audio and video with speech-to-text transcription and structured technical assessment summaries.
Normalised five-pillar matching
MatchNormalizer converts job requirements and candidate credentials into comparable schemas. MatchingAgent computes core-skill overlap, framework/tool coverage, experience, external evidence and learning velocity. The default weights are 35/20/20/15/10, with a separate job-specific multiplier on learning velocity. The response includes the weighted breakdown, matched/missing skills and decision enums, allowing the score to be traced back to its components.
Ed25519 credentials and live interview transport
Passport signing operates on canonical JSON and exposes an explicit verification endpoint. The signature binds the issued credential to the signing key and detects payload changes. The separate Gemini Live interview flow handles audio/video interaction, transcription and report generation. Candidate applications, company role pipelines and reviewer screens connect these services into the broader product workflow.
04 / BUILT WITH
TOOLS & THEIR PLACE- 01Python 3.11Backend, agents & orchestration
- 02Ed25519Signed credential payloads
- 03Canonical JSONDeterministic signing input
- 04Evidence graphProvenance & conflicts
- 05MatchNormalizerComparable scoring schemas
- 06SQLAlchemy / PostgreSQLResumable pipeline state
- 07ASGI / WSGIIn-process agent dispatch
- 08Gemini LiveRealtime interview sessions

