A
Candidate onboarding is labor-heavy
- Manual script delivery, room/desk/item checks
- Manual ID and documentation validation
- No consistent guided workflow for the candidate
AI-Enhanced Remote Proctoring · Proposal
Lyzr×Firstsource×ETS
Lower cost. Better experience. Full governance. The question is how to rebuild ETS's proctoring model — not whether.
Proposed by Lyzr + Firstsource
- Audience
- Head of Ops · CIO
- Deployment
- AWS VPC
- Term
- 3 Years
- Model
- AI + Human in the Loop
Case for Change
ETS's current proctoring model creates cost, experience, and competitive exposure.
Current-state pain points
Source: ETS-shared materials
B
Live monitoring is human-intensive
- Heavy reliance on human judgment with high cognitive load
- Variable consistency across sessions
- Limited ability to surface the right anomalies in real time
C
Post-exam QA is expensive and slow
- 6–8% of sessions re-reviewed manually
- Evidence gathering is fragmented
- Compliance review not consistently standardized
D
PSI Bridge transition creates risk and opportunity
- Migration work already underway
- ETS cannot absorb another high-touch transformation
- Any AI initiative must run as sidecar — minimal ETS lift
21M
exams/yr
Pearson VUE delivering with 2nd-gen AI sensors
73%
AI-assisted (2025)
US proctoring adoption is AI-assisted
$988M
market (2025)
Global remote proctoring market
Peer competitors are shipping AI proctoring at production scale. The market has converged on AI + Human models — ETS's current model is the outlier.
The Proctoring Inflection Point
ETS's measurement science is unmatched. The proctoring delivery model is the bottleneck — and every peer competitor has already moved.
The Solution
AI + Human in the Loop. Not AI-only. Not human-only.
AI handles repetitive checks, continuous monitoring, and QA triage. Humans supervise, adjudicate, and train the system. This is where market leaders are converging — and where ETS's proposed model is strategically strongest.
Design principles
01
Deploy in ETS cloud / VPC
ETS owns the runtime, storage, and data boundary.
02
Integrate, don't disrupt
Work in parallel with PSI Bridge, SORs, and related systems. No disruption to V-box migration.
03
Automate only what should be automated
Deterministic tasks and strong-signal detection go to AI. Ambiguity stays with humans.
04
Make every alert explainable
Every AI decision has trace, evidence, and rationale. No black-box flagging.
05
Use low-cost CV by default
MediaPipe and open-source vision for continuous monitoring. Commercial APIs (Rekognition) only for high-value checks.
06
Human feedback improves the system
Supervisor corrections, overrides, and adjudications become training inputs. The system learns from use.
Three-agent architecture
01Detection
Onboarding Assistant
Candidate greeting and script orchestration
Sponsor / exam rule retrieval
ID and liveness workflow orchestration
Environment verification and readiness summary
02Vigilance
Vigilance Core
Fuse CV detections with session context
Distinguish low-value anomalies from real risk
Prioritize alerts (low / medium / high)
Prepare intervention context for the proctor
03Adjudication
Forensic QA
Re-run checks on sampled sessions (6–8%)
Organize evidence by timestamp, issue, severity
Compare onboarding baseline vs. exam conditions
Produce QA report and disposition suggestions
Then → Human Proctor
Reviews AI alerts and intervenes when needed. Clears false positives, approves or rejects QA findings, and provides feedback that retrains the system.
What AI does vs. what humans do
AUTOMATED
AI handles
- Repetitive script handling and check-in flows
- First-pass validation and anomaly surfacing
- Confidence scoring and structured summarization
- Suggested actions and batch QA review
- Evidence assembly and audit preparation
SUPERVISED
Humans handle
- Ambiguous case review and intervention
- Policy exceptions and exam termination
- Adjudication and final QA approval
- Feedback correction and training input
AI does not make final punitive decisions. Alerts are reviewed by humans. The evidence path is transparent. Quality improves through reviewer feedback. The system is tunable by sponsor, program, and test type. This is why the HITL model is strategically stronger than either fully manual or fully automated proctoring — and why ProctorU/Meazure discontinued AI-only services in favor of this exact approach.
What Lyzr brings
Seven platform capabilities that make the three-agent architecture production-ready, governed, and continuously improving.
Govern & Trust
Responsible AI / Guardrails
Automated safety checks on every agent interaction — security, privacy, brand risk, content quality. Each check independently configurable.
Prevent the AI proctor from leaking test content, surfacing biased alerts, or making decisions outside its defined scope. Integrates with AWS Bedrock Guardrails.
GitAgent
Every agent lives in a git repo — identity, rules, memory, tools are version-controlled files. Every decision is a commit. Full audit trail out of the box.
Know exactly which agent version ran during any exam session. Roll back a bad update instantly. Compliance teams review agent config like a code PR — SOX, GDPR, FERPA aligned.
Observability
Real-time visibility into every agent action, following OpenTelemetry standard. Detailed per-step traces, aggregated overview, and downloadable reports.
Trace exactly how the Vigilance Core reached a flagging decision — which CV signal triggered, what confidence score, what context was evaluated. Every alert is explainable.
Test & Improve
Agent Evaluation (Simulation Engine)
Test agents against synthetic conversations before deployment. Scores across task completion, hallucination, faithfulness, toxicity, bias, and tool accuracy.
Before deploying to live exams, run the Onboarding Assistant against 1,000 synthetic check-in scenarios — edge cases, accessibility needs, unusual environments. Fix failures before candidates see them.
Agent Improvement Engine
Monitors live agents continuously, detects quality issues from production traces, and auto-generates prompt improvements to address them.
If the Forensic QA agent starts producing lower-quality evidence reports, the engine detects the drift, diagnoses the cause, and suggests a fix — before ETS's QA team notices.
Scale & Own
Control Plane
Centralized registry and management across all agents, all frameworks, all clouds. One view of your entire agent workforce — monitoring, guardrails, access controls, deployment.
One dashboard showing Onboarding, Vigilance, and Forensic QA agents across all exam programs. Performance, cost, alert volume, and compliance status — per program, per region.
ShadowLM
Progressively moves agent workloads from frontier models to small, owned models trained on ETS's approved decision traces. Up to 90% inference cost reduction.
Routine proctoring decisions — clear environments, standard ID matches, normal behavior — shift to a model ETS owns and runs in-VPC. Frontier handles edge cases only. Cost bends down permanently.
7
platform capabilities included
0
additional vendors to procure
100%
of agent actions traceable and explainable
Why this combination wins
ETS/PSI testing domain and platform ownership. Firstsource operating muscle and trained proctors. Lyzr agent orchestration, explainability, and deployability. A cloud-native, sponsor-aware, feedback-learning design — with minimal disruption to the current migration.
Delivery Roadmap
Lyzr + Firstsource own end-to-end delivery. ETS provides approvals.
The goal is zero disruption to the current PSI Bridge migration. Lyzr + Firstsource absorb the build burden. ETS involvement is limited to 1 weekly governance meeting, named approvers, 2–3 focused workshops, and UAT sign-off.
Phase 0
Mobilization & Access
2–3 weeks
- Named counterparts and architecture sign-off
- Access plan and deployment blueprint
- Sponsor rulebook and current process pack
- KPI baseline
Phase 1
Foundation Deployment
2–4 weeks
- Deployment in ETS cloud / VPC
- Logging, observability, security controls
- Connectivity to PSI Bridge, SORs, DBs
Phase 2
Onboarding MVP
4–6 weeks
- AI-guided check-in flow
- Liveness + identity verification
- Rule-based item/room scan
- Readiness summary and proctor handoff
Phase 3
Live Proctoring Build
4–6 weeks
- Anomaly detection and proctor alert console
- Threshold tuning and event trace
- Intervention and escalation workflows
Phase 4
Post-Exam QA
3–4 weeks
- Sample re-review engine (6–8% of sessions)
- Evidence package and audit-ready reports
- Issue classification and trend view
Phase 5
Pilot, Hardening & Enablement
4–6 weeks
- Pilot rollout with precision/recall tuning
- False-positive reduction
- Training, SOPs, runbooks for proctors/supervisors
Timeline
Total: 4–6 months to first scaled pilot
Lyzr + Firstsource own
- Solution deployment in ETS cloud
- API integration to PSI Bridge and internal systems
- Agent design, build, testing, hardening
- Prompt/rules/knowledge base setup
- UAT support, enablement, runbooks, hypercare
ETS provides
- Cloud/security approvals
- Access to named systems
- Sample rules, scripts, SOPs
- 1 product owner / business lead
- Periodic review and UAT sign-off
Architecture
A layer above the existing stack, not another tool inside it.
ETS keeps its LLM contracts, cloud infrastructure, and assessment systems. Lyzr sits between — orchestrating, governing, and auditing. Nothing replaced.
Agent Experiences
What ETS's candidates and staff experience
Live AI ProctoringIdentity VerificationEnvironment MonitorChat AI ProctorSession ReviewProgram Analytics
Lyzr Platform Layer
Lyzr Agentic Workbench — the control plane
Agent orchestration, identity verification, behavior detection, chat engine, guardrails, observability, ShadowLM, RBAC — all governed, all auditable, all in ETS's VPC.
Agent StudioMulti-Agent MeshVision PipelineIdentity AgentsBehavior ModelsChat EngineShadowLMGuardrailsObservabilityDecision InboxAudit Trail
ETS Infrastructure
What ETS already pays for — unchanged
Lyzr reads from and writes to these systems. It holds no canonical state.
AWS / AzureBedrock / OpenAIPSI BridgeAssessment PlatformScoring EngineCandidate DBContent Management
Technology pattern
Use low-cost computer vision by default. Use commercial APIs selectively.
MediaPipe + Open-Source CV
Continuous monitoring
- Face landmarks and gaze estimation
- Head pose and body posture
- Multiple-person detection
- Object emergence and item classification
- No per-frame commercial API cost
Amazon Rekognition
Selective, high-value
- Face liveness verification
- Face compare / identity persistence
- Optional spot verification events
- Priced per-check, used selectively not continuously
Data sovereignty
FEATURE
Single-tenant VPC, multi-region
Deployed inside ETS's AWS account with region-isolated instances. Candidate data generated in EU stays in EU. GDPR, CCPA, jurisdiction-specific requirements enforced at infrastructure level.
FEATURE
Private inference
ShadowLM models run in-VPC per region. Zero external API calls for candidate data.
FEATURE
FERPA · GDPR · SOC 2 · Biometric
Compliant with FERPA (US), GDPR (EU), SOC 2, and state biometric privacy laws (BIPA, CCPA). Data residency enforced per region — as architecture, not policy.
FEATURE
Calibrated autonomy
Assist → Copilot → Semi-autonomous → Autonomous. Set per exam program and per region. Enforced at runtime, recorded in immutable audit trail.
5-Year Cost Model
Unit economics normalize as volume scales. Lyzr stays flat.
Lyzr platform is flat at $1.5M/yr. LLM and AWS scale with volume. ShadowLM bends inference cost down. Per-exam cost drops 72% as volume grows nearly 5×.
$1.71
Per-exam cost, Year 1 · 1.2M exams
$0.48
Per-exam cost, Year 5 · 5.8M exams
72%
Unit cost reduction over 5 years
$2.68M
5-yr LLM savings via ShadowLM
Executive takeaway: total cost rises modestly with scale, but unit cost drops 72% by Y5 due to platform leverage, volume growth, and ShadowLM optimization.
Unit Cost per Exam Declines as Exam Volume Scales
Total annual cost remains controlled with ShadowLM while per-exam cost falls from $1.71 to $0.48 across Y1–Y5
Lyzr platformLLM with ShadowLMAWS infrastructurePer-exam cost
| Y1 1.2M | Y2 1.7M | Y3 2.6M | Y4 3.9M | Y5 5.8M | 5-Yr Total | |
|---|---|---|---|---|---|---|
| Exam volume | 1.2M | 1.7M | 2.6M | 3.9M | 5.8M | — |
| Lyzr platform (flat) | $1,500K | $1,500K | $1,500K | $1,500K | $1,500K | $7,500K |
| LLM (without ShadowLM) | $350K | $496K | $758K | $1,138K | $1,692K | $4,434K |
| LLM (with ShadowLM) | $350K | $347K | $379K | $341K | $338K | $1,755K |
| AWS infrastructure | $200K | $283K | $433K | $650K | $967K | $2,533K |
| Total with ShadowLM | $2,050K | $2,130K | $2,312K | $2,491K | $2,805K | $11,788K |
| Per-exam with ShadowLM | $1.71 | $1.25 | $0.89 | $0.64 | $0.48 | — |
| ShadowLM cumulative savings | $0 | $149K | $528K | $1,325K | $2,679K | $2,679K |
ShadowLM — a small model you own
The frontier model starts the work. ETS's model finishes it, and keeps it.
01
Frontier
Run the task. Log every run.
02
Distill
Fine-tune a small model on the logs.
03
Reinforce
RL until it matches the teacher.
04Optional
Mid-train
Inject domain knowledge.
Each rung entered only when the eval and the economics agree. Every human correction retrains the model. Currently being built with enterprise design partners.
Up to 90% lower cost
Routine proctoring decisions shift to owned model. Frontier handles edge cases only.
Data stays home
Model runs inside ETS's VPC. Candidate data never leaves the perimeter.
IP you own
Fine-tuned weights are an asset on ETS's balance sheet, not a subscription.
Volume grows nearly 5×. Total cost grows 37%. Per-exam cost drops 72%. ShadowLM saves $2.68M over 5 years.
What the alternative costs
Without Lyzr, ETS stitches together 9 vendor tools and 5 dedicated resources. The vendor stack is usage-based — costs scale with exam volume. Lyzr is flat. Server and LLM costs are common to both paths and excluded from this comparison.
Beyond licensing, 9 vendors means 9 NDAs, 9 security reviews, 9 SLAs, 9 integration maintenance cycles, and 9 renewal negotiations annually.
| Cost line | Y1 (1.2M) | Y2 (1.7M) | Y3 (2.6M) | Y4 (3.9M) | Y5 (5.8M) | 5-Yr Total |
|---|---|---|---|---|---|---|
| Software (9 vendors) | $1,255K | $1,770K | $2,460K | $3,575K | $5,180K | $14,240K |
| Team (5 resources) | $558K | $558K | $558K | $558K | $558K | $2,790K |
| Subtotal — what Lyzr replaces | $1,813K | $2,328K | $3,018K | $4,133K | $5,738K | $17,030K |
| Lyzr platform | $1,500K | $1,500K | $1,500K | $1,500K | $1,500K | $7,500K |
| Annual savings | $313K | $828K | $1,518K | $2,633K | $4,238K | $9,530K |
Server and LLM costs (~$550K/yr) are common to both paths and excluded. Total savings reflect only the components Lyzr replaces: software licensing and dedicated team.
View tool-level breakdown +
- LangChain + LangSmith: Y1 $400K → Y5 $1,800K (5-yr: $4,800K)
- Mem0: Y1 $250K → Y5 $1,100K (5-yr: $2,950K)
- Arize AI: Y1 $180K → Y5 $750K (5-yr: $2,030K)
- N8N: Y1 $100K → Y5 $400K (5-yr: $1,170K)
- Weaviate: Y1 $90K → Y5 $400K (5-yr: $1,060K)
- Guardrails AI: Y1 $80K → Y5 $350K (5-yr: $980K)
- Composio: Y1 $75K → Y5 $300K (5-yr: $850K)
- Lovable: $50K/yr flat (5-yr: $250K)
- Langfuse: $30K/yr flat (5-yr: $150K)
⚑ Pricing based on enterprise tier estimates scaled to exam volume. Revalidation in progress.
Annual cost of what Lyzr replaces — 9 vendors + team vs. Lyzr platform
$9.5M
5-year savings on software + team
3.8×
Y5 vendor cost vs. Lyzr platform
9 → 1
Vendor relationships eliminated
Note: the $9.5M reflects savings on software and team that Lyzr replaces. ShadowLM's $2.68M savings (shown above) are additional — reducing LLM inference cost within the Lyzr path. These are separate, non-overlapping benefits.
Assumptions
| Parameter | Value |
|---|---|
| Lyzr license | $1.5M/yr flat · 3-year term · includes build, no cap on agent actions |
| LLM base cost | $350K/yr at 1.2M exams · scales proportionally with volume |
| AWS base cost | $200K/yr at 1.2M exams · scales proportionally with volume |
| ShadowLM reduction | Y1: 0% · Y2: 30% · Y3: 50% · Y4: 70% · Y5: 80% |
| Volume model | 10% penetration by Y5 · Y3-Y5 growth ~50% YoY |
| Per-exam unit cost | (Lyzr + LLM + AWS) ÷ exam volume |
| Penetration benchmark | Enterprise AI handles 30% of customer interactions today (Zendesk/McKinsey). At 10% by Y5, ETS's model is conservative. |
- ⚑ ShadowLM reduction rates illustrative. Currently being built with enterprise design partners.
- ⚑ Per-exam cost excludes Firstsource human supervision/QA. Separate line item.
- ⚑ At 5.8M sessions, Lyzr flat pricing may warrant review. Flag for internal discussion.
- ⚑ Software pricing based on enterprise tier estimates. Revalidation in progress.
The Question
ETS's defensibility is measurement science, not platform engineering. Lyzr handles the platform. ETS ships the science.
Closing Directive
Next Steps
STEP
Discovery Workshop
Validate pain points, map sponsor rules, confirm integration architecture.
2 weeks
STEP
Pilot Scope
Select 1–2 exam programs, define success metrics, set autonomy levels.
2 weeks
STEP
Production
Deploy, test, harden, enable.
4–6 months to scaled pilot
Lyzr × Firstsource × ETS · AI-Enhanced Remote Proctoring Proposal · Confidential · June 2026