Screener

Khan Academy PM Content Operations

Screener Proposal

Victor Meinert | June 2026

While the brief calls for an intern-based workflow, this proposal presents a faster, more scalable, lower-cost alternative that uses AI to eliminate the production bottleneck entirely, and a working proof of concept to back it up.

This proposal recommends a solution that utilizes agentic agents to process all raw video content to automatically generate the required metadata/SEO. AI generates, and humans verify. The result is 1,000 videos published in under three months at a measurable cost savings and a new agentic pipeline for Khan Academy that can scale beyond 1k videos to 10k, or 100k.

A successful proof of concept has been completed and can be accessed at /poc.

Section 1

Workflow & Capacity Planning

Phase 1: Infrastructure & Foundation

Weeks 1-3PM owns

→Build, context, test, and harden the AI metadata agentic engine (recommend Claude)*
→Setup Jira: states, automations, Review Time Allowance (RTA), escalation triggers, reports, etc.
→SME Advisor board: Assemble knowledge base from SME documentation + platform standards
→Publish end-to-end pipeline documentation that covers responsibility assignment framework (RACI). Includes all stakeholders: IT/SME/QA/CMS/PM, + business owners.
→Stand up cloud storage, security, and asset access controls, CMS, etc.

Proprietary information and data will not be shared back to the AI provider.

MVP Gate

Phase 1 closes with a formal end-to-end content test: a 10% subset of actual Khan Academy raw videos run through the system and every element of the system is exercised, every stakeholder is actively engaged. This MVP cycle is the key mitigation mechanism for shielding SMEs from unnecessary or uncontrolled distraction during the live production phase and is expected to be the highest SME time commitment. SMEs from each subject area will provide hands-on review of MVP outputs and provide feedback and explicit approval before production begins. Iterate as necessary.

Because the processing solution uses the SMEs actual content (their words, their concepts already in the videos) there should be a high level of confidence that the outputs will be consistently contextually accurate and therefore high value.

Phase 2: Agentic Asset Processing

Fully Automated

All 1,000 videos enter the pipeline as a bulk run.

→Agent ingests each video and transcribes audio
→Generates 5 outputs: title, SEO description, timestamps, tags, thumbnail brief
→Deposits video + metadata to structured cloud storage
→Auto-creates a Jira ticket, links the asset, notifies QA team with status: Ready for Review

Estimated token cost for 1,000 videos in API fees: ~$50 and 8 days

Could also be staged if desired to reduce risk, but compute cost is low enough to not prioritise chunking.

Phase 3: QA Review

Human Gate

QA interns verify AI outputs. Each ticket has a built-in Review Time Allowance (RTA), the total budgeted time per asset covering initial review, defect logging, and re-review. Enforced natively in Jira.

→

Standard RTA: 20 minutes per video

→

Exception ceiling: 30 minutes for complex or high-defect videos

→

Remediation: AI agent re-generates flagged outputs, QA reviewer re-reviews (does not rewrite)

→

SME spot check: SME spotchecks outputs for content accuracy patterns

→

Full escalation: triggered only when data hallucinations occur or content is factually ambiguous beyond the knowledge base

Extra validation could be achieved with agentic agents running SEO routines against CMS outputs.

Phase 4: Agentic Publishing

Fully Automated

Approved tickets trigger the publishing agent. Video and metadata push to the Khan Academy CMS. Ticket closes. PM maintains a notification bridge so CMS stakeholders know what's inbound.

Total System Time Per Video

~4 minMetadata processing and outputautomated

20 minQA review (RTA), human-only phasehuman only

~1 minPublishingautomated

Total: ~30 minutes per video. 20 minutes is human. All else automated.

Section 2

Resource Pitch & Budgeting

Because interns verify rather than produce, 3 reviewers are sufficient to cover 1,000 videos in under three months with significant capacity to spare. Infrastructure for AI compute may require IT support and scalable (AWS) cloud infrastructure so cost there is tbd but should be minimal. Cost for AI infrastructure can be amortized across future projects.

→3 India-based QA reviewers

→Stipend: 17,500 INR/month (~$184 USD), current India market rate for content ops / QA roles

→Duration: under 3 months

Total QA pod cost: ~$1,800 USD | AI API cost: ~$50 | Total: ~$2,000

Capacity

If we assume the project is required to reach 100% completion in exactly 3 months and the system architecture consumes 3 weeks of that budget, this proposal targets a 6 week QA phase.

→Intern RTA = 20 min/video

→3 QA Interns x 30 billable hrs/wk = 90 hrs/wk

→Throughput Potential = 270 videos/wk

→6wk Capacity = ~1,620 videos

Capacity buffer to be applied to initial training/ramp-up and defect loops, re-reviews.

90 hours x 60 min = 5,400 min divided by 20 min per video = 270 videos/week

SME Time Commitment

Full transparency on what this model asks of SME stakeholders:

→Knowledge base + agent training: 1 day per SME, one time (Phase 1)

→MVP Gate review and approval: ~2 days per SME, one time (Phase 1)

→Spot checks every 10th video: ~5 min x 100 checks = 1 day total (Phase 3) (tbd necessary)

→Full escalations: volume tbd, calibrated after MVP Gate

Total per SME controlled capacity commitment: ~3-4 days across 3 months.