The classroom video-collection & AI-analysis platform: teachers upload lesson videos from the web or Telegram, the system transcribes Khmer/English audio, corrects it with an LLM, scores it against Early Grade Learning teaching practices, and routes it to a staff review workflow.
Production: egl.openplp.com · VPS 209.74.72.138 · Status & limitations review · July 2026
Early-grade teaching quality across Cambodian provinces can't be observed at scale. In-person observation is expensive, slow, and reaches a handful of classrooms.
Almost no commercial AI service transcribes Khmer classroom audio reliably. Off-the-shelf tools built for English don't transfer.
Many teachers work primarily from phones with Telegram, not laptops with browsers. Collection has to meet them where they are.
The bet: combine self-hosted Whisper transcription + LLM post-correction + Claude-driven EGL rubric analysis so every uploaded lesson becomes reviewable, searchable evidence of teaching practice.
Stack: TanStack Start SSR (Bun, shadcn/ui, Tailwind v4) · FastAPI + SQLAlchemy 2.0 + PostgreSQL · faster-whisper · Claude tool-use analysis.
/ → SSR frontend, /api → FastAPI,
/api/uploads → video files.
_authenticated/, JWT in localStorage,
dark/light theming, Framer Motion.
/var/egl/uploads,
served back via nginx.
Email/password · Google Sign-In · Telegram Login Widget · Cloudflare Turnstile bot-check gate · SMTP password recovery. First user created becomes super_admin.
No API keys in .env. Super-admin settings UI writes
AES-256-GCM-encrypted credentials to per-provider tables; secrets are write-only
over the wire, every change audited.
A local telegram-bot-api server lifts the cloud 20 MB download cap
to ~2 GB so real classroom videos can flow in through the bot.
Filterable list with processing badges; global search over title/subject/notes (staff see everything, teachers see their own).
Blanket consent covers all existing + future videos; per-video withdrawal supported; state persists across reloads and shows on every video detail page.
Personal stats page plus static recording guidelines so teachers know what a good lesson video looks like.
/start <code> to the bot.
Why it matters: this is the highest-leverage adoption feature — a teacher with only a phone and Telegram can contribute a fully tagged, consented, AI-analyzed lesson video without ever opening the web app. The same pipeline, notifications, and review workflow apply regardless of source.
ffmpeg transcodes HD → 720p H.264 MP4 (same UUID, URL stays valid) — only kept if smaller. A storage/playback win; transcription cost is set by audio duration, not resolution.
faster-whisper large-v3 (best available for Khmer). Audio decoded to 16 kHz mono once, then two passes: native language (auto-detected) + English translation.
Post-ASR LLM pass fixes Whisper's systematic Khmer errors (homophones, word segmentation, numerals). Timestamps preserved; raw output kept for audit; skipped for English audio.
Claude (tool-use, structured output) scores the lesson 0–100 and extracts EGL practices — modelling, checking understanding, wait time, feedback, scaffolding, student talk — each with evidence + timestamp.
Designed to degrade gracefully: every stage is optional — missing ffmpeg, missing Whisper deps, or no LLM key skips that stage rather than failing the video. Failures notify the teacher in-app and via Telegram, and staff can hit Reprocess on any video.
raw_segments — corrections are
auditable, never destructive
Staff see a 30-day upload trend, geographic breakdowns by province/school (inherited from teacher MSS profiles), teacher coverage, and review outcomes (approved / needs work / rejected / pending).
One review per reviewer per video: approve / needs-work / reject, optional 1–5 rating + comment. Teacher gets a named notification on every decision. Review queue lives in the admin videos grid.
Role changes (teacher ⇄ reviewer ⇄ admin ⇄ super_admin) with safeguards: no self-escalation, the last super_admin can't be demoted. Linked identities (Google/Telegram) visible per user.
Tabbed provider configuration — Anthropic, OpenRouter, Mistral, Google, Telegram, Turnstile, SMTP, PLP MSS — all encrypted at rest, write-only over the wire, with per-provider “test” actions.
One-click background sync pulls the full MSS hierarchy (zone → school, ~13 API calls), idempotent upserts, status polling in the UI, cron-callable endpoint for automation.
Super-admin can trigger reminder emails to teachers with videos still in the pipeline; endpoint is cron-callable for scheduled nudging later.
Zone → province → district → commune → cluster → school, pulled from the MSS API with encrypted credentials. Drives the profile school picker; every video inherits its uploader's geography, powering the province/school dashboards.
Gap: MSS has no villages — the geo_villages table exists but
stays empty.
Complete national tree — 25 provinces, 197 districts, 1,646 communes, 14,372 villages, bilingual Khmer + Latin — seeded from a static JSON file with no network or credentials needed.
Why not merged: the datasets share only ~70% of names / ~80% of school codes — welding them would orphan data, so they intentionally power two independent pickers.
| Capability | Status |
|---|---|
| Web + Telegram video upload | ✓ shipped |
| Khmer/English transcription + translation | ✓ shipped |
| LLM transcript correction | ✓ shipped |
| EGL analysis (score + practices) | ✓ shipped |
| Consent (blanket + per-video) | ✓ shipped |
| Notifications (in-app + Telegram + email) | ✓ shipped |
| Admin: users, videos, settings, stats | ✓ shipped |
| Geographic sync + gazetteer | ✓ shipped |
| Search | ✓ shipped |
| Capability | Status |
|---|---|
| Dedicated reviewer UI (submit reviews) | ◐ backend full, frontend partial |
| Khmer-language UI | ✗ English-only (Khmer in bot only) |
| CSV/XLSX export | ✗ planned |
| In-browser recording | ✗ planned |
| Live transcript (websocket) | ✗ planned |
| Audit-log viewer UI | ✗ table exists, no UI |
| 2FA | ✗ not implemented |
| Cloud storage / CDN | ✗ local disk only |
| Persistent job queue | ✗ in-process threads |
Practical ceiling: the current architecture comfortably serves a pilot (tens of teachers, a few uploads/day). A provincial or national rollout requires object storage, a real job queue, a GPU (or hosted ASR) for transcription, and a backup/restore story — in that order.
Khmer transcription accuracy is bounded by Whisper large-v3 — the best option available, but still a low-resource-language model. Noisy classrooms, young children's voices, and chorus responses degrade it further. No commercial provider currently offers a better Khmer alternative.
The chain is transcript → translation → Claude analysis. An ASR mistake survives correction, skews the English translation, and can shift the EGL score. Scores should be treated as a triage signal for human reviewers, not a measurement instrument.
Every video incurs LLM cost (correction + analysis) plus heavy CPU for Whisper. Analysis quality depends on external API availability and keys configured by the admin; with none configured, videos are collected but never analyzed — the platform silently becomes a plain video locker.
No ground truth yet: there is no human-verified transcript or scoring benchmark in the system, so ASR word-error rate and analysis accuracy for this specific corpus are unmeasured. A small gold-standard set (20–30 human-reviewed videos) would make the quality claims testable.
The bottleneck is transcription: large-v3 on CPU with beam 5 costs roughly 1–2× the video's duration on 16 modern cores — a 40-minute lesson ties up a worker for 40–80 minutes. Everything else (RAM, transcode, LLM calls) is comfortably cheap.
| Volume | CPU | RAM | Storage | Network | Est. price / mo |
|---|---|---|---|---|---|
| Pilot · ≤20 videos/day (~500/mo) | 16 cores (dual Xeon Silver 4208 or equiv.) | 64 GB | 2× 1 TB NVMe | 1 Gbps | $228–241 |
| Bulk · 50–100/day (~1,500–3,000/mo) | 32 cores (dual Xeon Gold 5218 / EPYC 7313), or 16 cores + GPU ≥16 GB | 128 GB | 4× 1–2 TB NVMe | 1 Gbps unmetered | $285–302 |
| Heavy · 100+/day | GPU server (L4 · A4000 · RTX 4090) — Whisper runs 10–20× realtime | 64–128 GB | 4 TB+ NVMe | 1 Gbps unmetered | $300–450 |
Sizing notes: plan ~0.6 GB stored per video post-transcode (+30% headroom,
originals briefly double it) · uploads cap at 2 GB each, so 1,000 videos/mo ≈ 1.5 TB
ingress — pick unmetered or ≥20 TB traffic · a GPU is cheaper per video than more
CPU cores at bulk volume, and lets processing_workers scale past 2.
Pilot/Bulk prices are Namecheap dedicated-server list rates billed month-to-month
with no annual commitment
(not the discounted 1st-year rate); Heavy is a rough estimate pending a quote.
| Scale | Videos / mo | Dedicated server (est. / mo) | Claude API (standard) | Claude API (Batch −50%) | Total / mo (est.) | Package price (+50% margin) |
|---|---|---|---|---|---|---|
| Pilot | 500 | 16-core · 64 GB · 2×1 TB NVMe — $228–241 | ~$285 | ~$143 | $371 – $526 | $557 – $789 |
| Bulk | 1,500 | 32-core · 128 GB · 4×1–2 TB NVMe — $285–302 | ~$855 | ~$430 | $715 – $1,157 | $1,073 – $1,736 |
| Heavy | 3,000 | GPU server (L4 / A4000 / 4090) — $300–450 | ~$1,710 | ~$855 | $1,155 – $2,160 | $1,733 – $3,240 |
| National | 5,000 | GPU server + storage growth — $400–550 | ~$2,850 | ~$1,425 | $1,825 – $3,400 | $2,738 – $5,100 |
How to read the total: low end = Batch API + cheaper server tier; high end = standard API + pricier server tier. Pilot/Bulk server costs are Namecheap month-to-month dedicated pricing (no annual prepay discount); Heavy/National remain rough Hetzner/OVH-class estimates pending a quote. Assumes 40-min Khmer lessons — English audio costs ~⅓, shorter lessons scale down linearly. The server price is fixed; the Claude API line is the only cost that grows with volume. Package price = Total cost × 1.5 (50% margin) — a starting point for what to charge as a monthly subscription, not a final quote.
Guiding principle: the platform's collection and workflow layers are genuinely strong; its risks are concentrated in infrastructure durability and unvalidated AI quality. Spend first on not losing data, then on the people using it, then on scale.
End-to-end teacher flow on web + Telegram, a genuinely rare Khmer AI pipeline with auditable corrections, encrypted-by-design integrations, and national geographic data built in.
Single-VPS, local-disk, thread-pool architecture; Khmer ASR at the mercy of Whisper's ceiling; English-only UI; review loop and exports unfinished.
Treat AI scores as reviewer triage, invest in backups + Khmer UI before scaling, and build a gold-standard set so quality becomes measurable instead of assumed.
EGL Learning Labs · egl.openplp.com · generated from a full codebase review, July 2026