HireFinch

HireFinch: Voice-Based AI Interviewing at Enterprise Scale

Executive Summary

HireFinch is Techjays’ voice-based interviewing system that conducts structured interviews, scores candidates against job-specific rubrics, and generates evidence-backed summaries within minutes. It combines a provider-agnostic inference router (OpenAI Realtime, Gemini Realtime, and a cost-optimized TTS tier), agentic hiring analytics, proctoring signals for integrity, and privacy-by-design controls.

Business outcomes observed in pilots

Human screening load reduced by approximately 96 percent From 5000 to 10000 initial screens per 100 roles to 200-400 finalists
Interview report turnaround: manual to minutes
Training and education: more than 90 percent cost reduction with TTS tier, maintaining acceptable conversation quality
Multi-provider failover ensures service continuity during regional or provider issues

Engineering guarantees

Availability SLO: 99.95 percent monthly uptime
p95 reply time: ≤ 1.2 seconds on realtime tier
ASR quality: Word Error Rate ≤ 8 percent at p50 and ≤ 15 percent at p95
Explainability and fairness:
Rubric-adherence F1 ≥ 0.85
Counterfactual stability ≥ 95 percent across accents, devices, and noise environment

Problem Context

High-volume hiring creates tension between thorough evaluation and operational efficiency. Teams either:

1. Over-index on resumes, introducing bias and missing true capability, or

2. Spend excessive interviewer time on first-round screening

This slows hiring, impacts candidate experience, and increases subjectivity.

Voice interviews offer stronger assessment of communication and reasoning but require:

Realtime reliability
Consistent scoring
Proctoring safeguards
Compliance-ready auditability

HireFinch addresses these needs at enterprise scale.

Solution Overview

HireFinch delivers structured, rubric-grounded voice interviews with explainable scoring and agentic analytics. It is deployed as microservices with event-driven orchestration, multi-region availability, and provider failover (OpenAI, Gemini, TTS tier) to optimize cost, latency, and reliability.

High-Level Architecture

Figure 1: High-Level Architecture

Legend: WebRTC = Web Real-Time Communication, ASR = Automatic Speech Recognition, VAD = Voice Activity Detection, EOT = End-of-Turn Detection, ATS = Applicant Tracking System.

Interview Flow (Realtime)

Figure 2: Interview Flow (Realtime)

HireFinch conducts a dynamic interview, analyzing responses turn-by-turn and generating insights immediately after completion.

Core Capabilities

Rubric-Grounded Scoring and Explainability

Every score cites transcript spans and rubric criteria
Calibration maintained against human panels (expected calibration error ≤ 0.05)
Rubric and skill definitions version-controlled with full audit diffs

Agentic “Talk-to-Data” Analytics

Talent teams ask open-ended queries to analyze cohorts
Executive summaries used first; full transcripts referenced when deeper comparison is required
Scales efficiently with large candidate volumes

Progressive Refinement with Human-in-the-Loop

When human reviewers reject candidates or add notes, the system updates:
               1. Skill taxonomy
               2. Missing questions
               3. Rubric thresholds
All changes are traceable and reversible

Integrity and Proctoring
‍

Multi-signal fusion including stylometry, lexical diversity, latency patterns, prosody shifts, periodic webcam snapshots, and deepfake detection features
False positive rate maintained ≤ 2 percent at 95 percent recall
Webcam images auto-deleted within 7 days; only derived descriptions retained

Privacy, Security, and Governance

Privacy-by-Design

Avoids prompting for unnecessary PII
PII-scrubbed content used in model operations

Encryption and Access Controls

Envelope encryption with KMS or Bring-Your-Own-Key support
Row-level access control and region-specific data residency (United States or EU)

Compliance and Audit

SOC 2 and ISO-aligned operational controls
Immutable audit logs and incident response playbooks

Data Model (PII Minimization)

Figure 3: PII-Minimized Data Model

Reliability, Cost, and Observability

SLOs and Failover

99.95 percent availability target
Recovery point objective ≤ 15 minutes
Recovery time objective ≤ 15 minutes for control plane

Failover uses normalized outputs to maintain consistent behavior during provider switching.

Figure 4: Provider Failover Lifecycle

Cost Optimization

Dynamic routing: simpler turns served by smaller models
Aggressive caching of transcripts and retrieval results
TTS tier reduces cost for academia and training scenarios

Observability

Per-turn traces include ASR metrics, model routing, retrieval context, scoring evidence, proctoring signals, and spend attribution
Burn-rate alerts enforce quality and cost safeguards

Evaluation Framework (Evals)

Figure 5: Evaluation and Release Gate Workflow

Test Datasets

Golden datasets: Diverse accents, locales, noise conditions
Counterfactual sets: Same answers rendered differently to ensure fairness stability
Adversarial sets: Deepfake voices, screen-reading, and prompt injection attempts

Release Gates

Speech and UX: WER thresholds, turn latency, barge-in, TTS MOS ≥ 4.2
Scoring: F1 ≥ 0.85 and evidence recall ≥ 90 percent
Fairness: Score parity within 2 percentage points
Safety: Zero PII leakage violations

Continuous integration enforces gates on every release with shadow evaluations and auto-rollback.

Deployment and Integration

Native ATS and email system integrations with retries and compensation flows
Minutes-based billing aligned to interview duration
Multi-tenant isolation with dedicated keys and rate-control fences

Risks and Mitigations

Risk	Mitigation
Provider outage or regional degradation	Circuit breaker failover, warm standbys, normalized output
Accent or noise bias	Counterfactual testing, dataset expansion, calibration dashboards
Proctor false positives	Reviewer queue, FPR ≤ 2 percent
Prompt injection or policy abus	Input and output guards, red-team suites
Cost regression	Routing and caching strategies, cost monitoring alerts
Data exposure	7-day media retention limit, encryption, strict ACLs

Conclusion

HireFinch provides efficient, auditable, and fair AI-assisted interviewing. By combining structured scoring, agentic analytics, proctoring safeguards, and enterprise governance, it reduces screening effort by approximately 96 percent, accelerates hiring decisions, and maintains consistent quality across volume. SLO-backed reliability and continuous evaluation ensure performance at scale as models evolve.

Appendix A: Acronyms

ASR, ATS, BYOK, CI, COGS, DLQ, ECE, EOT, EU, FPR, HITL, ISO, KMS, LLM, MOS, PCM, PII, p50/p90/p95, RPO/RTO, SaaS, SLA/SLO, SOC 2, ROC, TPR, TTS, UX, VAD, WER, WebRTC(All expanded at first use in the document.)

Appendix B: Example Guard Policies

Do not prompt for unnecessary PII
Evidence required for all scoring decisions
Context-budget limits; expand scope only when needed
Output filtered for PII leakage and compliance violations

Appendix C: Sample KPI Dashboard Metrics

Latency: p50 and p95 response time
Speech quality: WER
Scoring quality: rubric adherence and attribution
Fairness: parity deltas and counterfactual stability
Proctoring: flagged-turn rate and AUC
Reliability: uptime and failover events
Cost: spend per interview minute and cache hit rate

Techjays is committed to responsible AI adoption. HireFinch is engineered to deliver measurable hiring value without compromising privacy, fairness, or reliability.

Hi there! I am TEJA