Evalzo — Video Assessment Agent

Architecture Comparison

What Recall.ai Eliminated

What We Still Build

Three Approaches Compared (60-min interview)

Approach A: Naive — Full Video to Frontier Model

Approach B: Triage-First (v1 Architecture)

Approach C: Recall.ai + Groq Whisper + Assessment LLM (v2.1 Architecture)

Why Is v2.1 So Much Cheaper?

1. No triage LLM at all. v1 spent $0.40–0.80 per interview having a model watch the full video just to classify segments. Recall.ai gives us screenshare_on/off events for free as structured data.

2. No diarization needed. Recall.ai captures separate audio per participant (audio_separate_mp3). Each file IS a specific speaker — speaker identity is structural, not algorithmic. No pyannote, no GPU, no Celery workers.

3. Groq Whisper at $0.04/hr is 73% cheaper than Recall.ai's transcription. We skip Recall.ai's $0.15/hr transcription add-on and transcribe each participant's audio independently via Groq. 216x real-time speed means a 60-min call transcribes in ~17 seconds. And we own this layer — can swap to Deepgram or AssemblyAI if Groq pricing changes.

4. Screen-share video comes pre-separated. We don't need FFmpeg clipping. Recall.ai captures screen-share as a separate stream (type: "screenshare"). We send only that to the assessment LLM — no wasted tokens on webcam footage.

5. Lower frame rate is fine. For screen-share assessment (code on screen), we can sample at 0.5fps (one frame every 2 seconds). Code doesn't change that fast. This halves the token cost vs 1fps.

Dev Effort Breakdown — 1 Engineer, Full-Time

Assumes one backend engineer familiar with Python, REST APIs, and LLM integrations. All estimates include testing.

Implementation Guide — ClickUp-Ready

Detailed technical steps for developer handoff. Each section maps to a ClickUp task.

Task 1: Recall.ai Account + API Setup

Assignee: Backend engineer  |  Estimate: 2 hrs  |  Priority: Urgent

• Sign up at recall.ai/signup — free credits included for testing
• Get API key from dashboard. Region: us-east-1 (or eu-central-1 if EU data residency needed)
• Configure webhook URL in Recall dashboard for lifecycle events: recording.done, transcript.done, video_separate.done
• Store API key in Evalzo secrets manager (not in code)
• Test: curl -H "Authorization: Token YOUR_KEY" https://us-east-1.recall.ai/api/v1/bot/

Task 2: Create Bot Service

Assignee: Backend engineer  |  Estimate: 8 hrs  |  Priority: High

Service that creates a Recall.ai bot when an interview is scheduled.

POST https://us-east-1.recall.ai/api/v1/bot/
{
  "meeting_url": "https://zoom.us/j/123456789?pwd=abc",
  "bot_name": "Evalzo Assessment",
  "join_at": "2026-04-25T14:00:00Z",   // scheduled >10min ahead = guaranteed
  "recording_config": {
    // NO transcript provider — we use Groq Whisper on per-participant audio
    "audio_separate_mp3": {},            // per-participant audio (natural diarization)
    "video_mixed_mp4": {},               // 720p gallery view recording
    "video_separate_mp4": {},            // separate webcam + screenshare per participant
    "participant_events": {},            // join/leave/screenshare_on/off events
    "meeting_metadata": {}
  },
  "variant": { "zoom": "web_4_core" },  // required for video_separate_mp4
  "metadata": {
    "evalzo_interview_id": "int_abc123",
    "evalzo_jd_id": "jd_xyz456"
  }
}

• Store returned bot_id against the interview record in Evalzo DB
• Handle 507 (no capacity) with retry after 5s
• web_4_core variant adds $0.10/hr but is required for separate video capture

Task 3: Webhook Handler + Groq Transcription

Assignee: Backend engineer  |  Estimate: 8 hrs  |  Priority: High

Endpoint that receives Recall.ai lifecycle webhooks, transcribes via Groq, and triggers the assessment pipeline.

• Listen for recording.done event — this fires when all artifacts are ready
• On receive: fetch bot details via GET /api/v1/bot/{bot_id}/
• Extract from response:
  • Per-participant audio: GET /api/v1/audio_separate?recording_id={id} → download each MP3
  • recordings[0].media_shortcuts.participant_events.data.download_url → events JSON
  • Separate video: GET /api/v1/video_separate?recording_id={id} → filter for type: "screenshare"
• Transcribe via Groq Whisper V3 Turbo:
  • Send each participant's MP3 to POST https://api.groq.com/openai/v1/audio/transcriptions (model: whisper-large-v3-turbo)
  • Run in parallel — all participants transcribe concurrently (~17s total)
  • Each response includes timestamped words — tag with participant name from Recall.ai metadata
  • Merge all transcripts chronologically → fully speaker-attributed transcript, no diarization step
• Parse participant_events for screenshare_on / screenshare_off — extract timestamps
• Download screen-share MP4(s) to temp storage
• Enqueue assessment job with: merged transcript + screen-share video paths + JD ID

Task 4: Assessment LLM Service

Assignee: Backend engineer  |  Estimate: 16 hrs  |  Priority: High

Core intelligence layer. Consumes Recall.ai data + JD, outputs structured assessment.

• Model abstraction: Define interface assess(transcript, media, jd, config) → Assessment. Implement for Claude (via Anthropic API) and one backup (OpenAI or Google). Config file selects which.
• Frame extraction: Use FFmpeg to sample frames from screen-share MP4 at 0.5fps. Convert to base64 for multimodal API calls.
• Prompt structure:
  • System: Soul file + evaluation rules + output schema
  • User: JD (structured) + interview config + chronological interleave of transcript segments and video frames
  • Use JSON mode / structured output for consistent scoring format
• Two-pass option: Pass 1 = raw assessment. Pass 2 = JD alignment scoring. Can be same model, two calls.
• Context management: If transcript + frames exceed context window, chunk by interview section and run final synthesis pass.
• Cost tracking: Log input/output tokens + model used per assessment. Store in DB for monitoring.

Task 5: Output Formatter + Dashboard Integration

Assignee: Backend + Frontend engineer  |  Estimate: 12 hrs  |  Priority: Medium

• Transform LLM structured output into recruiter-facing format
• Generate evidence links: map transcript timestamps to Recall.ai video playback URLs
• Integrate with existing Evalzo candidate profile page
• Display: overall score, per-criterion scores, narrative summary, clickable evidence moments
• Show alongside QGen written test results for holistic candidate view

Task 6: Testing + QA

Assignee: Backend engineer  |  Estimate: 12 hrs  |  Priority: High

• Record 3–5 test interviews (mix of technical + behavioral + screen-share)
• Run through full pipeline, verify output quality
• Test edge cases: no screen-share, 90+ min interview, multiple screen-shares, bad network quality
• Benchmark: AI assessment vs human evaluator on same interviews
• Verify cost per assessment matches $1.65–3.65 projection
• Load test: 10 concurrent assessments

Recall.ai API Quick Reference

API Docs: docs.recall.ai/docs/getting-started  |  Rate limit: 120 requests/min  |  Auth: Authorization: Token {API_KEY}