DeepSight is an AI-powered YouTube/TikTok video analysis SaaS, built by Maxime Robathe (DeepSight Synthesis SAS, France) for European users who need source-grounded, nuanced video synthesis with privacy compliance.
This document explains how the stack is built, why we made specific technical choices (Mistral, Hetzner, Vercel), and how DeepSight differs from a "wrapper" around any single LLM API.
┌──────────────────────────────────────────────────────────────┐
│ USER SURFACES (TypeScript) │
├──────────────────────────────────────────────────────────────┤
│ • Chrome Extension v2.0 (Webpack 5, MV3) │
│ • Mobile App (React Native + Expo SDK 54) │
│ • Web App (React 18 + Vite 5) │
└──────────────────────────────────────────────────────────────┘
│
HTTPS + JWT Bearer
▼
┌──────────────────────────────────────────────────────────────┐
│ BACKEND (Python 3.11) │
│ FastAPI async, 4 workers, port 8080 │
├──────────────────────────────────────────────────────────────┤
│ 15 routers: auth, videos, chat, billing, playlists, study, │
│ academic, debate, history, admin, analytics, batch, │
│ notifications, search, transcripts │
└──────────────────────────────────────────────────────────────┘
│ │ │ │
▼ ▼ ▼ ▼
PostgreSQL 17 Redis 7 Mistral AI Perplexity AI
(Hetzner DE) (Hetzner DE) (France EU) (US — only for fact-check)
All compute and data live in the European Economic Area (Hetzner Falkenstein Germany + Mistral France). The only US dependency is Perplexity (used for optional fact-check enrichment when user activates web search) — which is explicitly opt-in per user/per query.
The honest answer: DeepSight uses Mistral as its LLM, but the value of DeepSight is in everything around Mistral, not the LLM call itself.
-
Multi-method transcript extraction (the hardest problem) — 7 methods in 3 phases:
- Phase 0: Supadata API (paid, reliable, our priority)
- Phase 1 (parallel): youtube-transcript-api + 10 Invidious instances + 8 Piped
- Phase 2 (sequential): yt-dlp manual subtitles + auto-captions
- Phase 3 (audio STT fallback): Groq Whisper → OpenAI Whisper → Deepgram → AssemblyAI
With circuit breaker, exponential backoff, health-check instance manager, user-agent rotation, DB cache L2 (chunked + embeddings), 12+ language support. No single LLM API gives you reliable YouTube transcripts.
-
Hierarchical chunking and analysis pipeline (v6) — videos are not uniformly long. Our pipeline routes content through 6 duration tiers (sub-15min → 15-30 → 30-60 → 60-120 → 120-240 → 240+) with adaptive chunking strategies. A 4-hour conference doesn't get analyzed like a 5-min short. This is product engineering, not LLM API.
-
Chat v4.0 with optional Perplexity enrichment — context-aware: detects when web search is genuinely needed (current events, contradicting claims) vs when video transcript alone suffices. Saves cost and latency when not needed.
-
Compliance and data sovereignty layer — EU hosting (Hetzner Germany), Mistral France LLM, encryption at rest (LUKS) and in transit (TLS 1.3), GDPR Art 22 disclosure, subprocessor list (see
/trust), DPA (Clauses Contractuelles Type EU 2021/914). This is not what a "wrapper" does. -
Domain-specific products built on the analysis primitive:
- AI Debate (compare 2 videos with opposition + complement + nuance)
- Quick Voice Call (talk to a video via ElevenLabs voice)
- Semantic search across user history
- Mind Maps (Cytoscape graphs)
- Spaced repetition flashcards
- Academic paper search linking (arXiv, Crossref, Semantic Scholar, OpenAlex)
A wrapper passes prompts to an API. DeepSight is a video intelligence platform that uses Mistral as one component.
We chose Mistral for 4 reasons, in priority order:
Mistral AI is a French company, GDPR-native, with EU data residency by default. For our target users (consultants in EU regulated sectors, researchers, privacy-conscious creators), this is not a "nice-to-have" — it's a binding constraint that disqualifies any US-hosted LLM API.
Concretely: DeepSight customers in banking, insurance, legal, public sector cannot use a tool routing prompts to OpenAI US under their internal compliance policies (DORA, NIS2, sectoral regulations).
Mistral 2026 catalog (as of May 2026):
mistral-small-2603— Free tier, fast, sufficient for short videosmistral-medium-2508— Pro tier, good quality/cost ratiomistral-large-2512— Expert tier, 262K context for long videosmagistral-medium-2509— reasoning model (used for AI Debate)voxtral-mini-2602— audio model (used for Quick Voice Call)mistral-embedv23.12 — embeddings for Semantic Search
We use the right model for each task. A 5-min YouTube short doesn't need the same model as a 4-hour conference, and AI Debate reasoning needs a different model than chat enrichment.
Mistral's pricing on mistral-medium is competitive enough to support our 8.99 €/month Pro tier with healthy margin. OpenAI gpt-4o-mini is also viable but the EU data residency would still be unsolved.
Mistral has European business hours, French support, and an active roadmap (Voxtral, Magistral, embeddings). We have a direct relationship that an OpenAI-via-Stripe SaaS doesn't give us.
Three reasons:
- Coherence of compliance promise — if we offer "OpenAI fallback" we break the "your data stays in Europe" promise. The promise is the product for a chunk of our market.
- Maintenance burden — supporting two LLM providers means doubling our test surface, prompt tuning, model-specific edge cases.
- Honest positioning — we're not trying to be the lowest common denominator. We're explicitly the European AI YouTube tool.
If a user requires OpenAI specifically, we recommend they use ChatGPT Plus directly. We're not for everyone.
That said, our LLM layer is OpenAI SDK compatible (Mistral exposes an OpenAI-format API). If a future Enterprise tier requires customer-side OpenAI keys for legal reasons, the integration cost is hours, not weeks.
- EU data residency — Hetzner is German, no Schrems II / Cloud Act concerns. AWS Frankfurt is technically EU but the parent company is subject to US legal compulsion.
- Cost — Hetzner is 5-10× cheaper than AWS for equivalent compute, which lets us keep the 8.99 €/month price point sustainable.
- Simplicity — bare metal-equivalent VPS, Docker stack, Caddy reverse proxy. No vendor lock-in. Migration to OVH or Scaleway possible in days if Hetzner ever fails us.
Tradeoff: less managed services. We run our own PostgreSQL 17, Redis 7, Caddy reverse proxy, monitoring. This is acceptable engineering cost for the price/sovereignty equation.
Vercel for the React/Vite web app, because:
- Instant global CDN — we want the marketing landing fast worldwide
- Atomic deployments — git push → preview URL → promote to prod in minutes
- Zero infrastructure management for static + edge-compute frontend
We don't host backend on Vercel — Functions cold-start and execution limits don't fit our long-running async tasks (4-hour video analysis).
1. User pastes URL on web/mobile/extension
POST /api/videos/analyze {url, mode, model}
↓
2. Cache check (L1 Redis + L2 Postgres) — if hit, return immediately
↓ (miss)
3. Transcript extraction — 7-method chain with circuit breakers
↓
4. Duration-based routing — chunk size & strategy per video length
↓
5. Mistral analysis — model selected by user's plan + analysis depth
↓
6. Optional fact-check via Perplexity (if user enabled web search)
↓
7. Result persistence — Postgres + Redis cache + analytics events
↓
8. SSE streaming to frontend with progress updates
End-to-end target: 30s for a 5-min video, 2 minutes for a 1-hour video, 8 minutes for a 4-hour conference.
Open (this document, GitHub repo):
- Architecture
- Pipeline design and tradeoffs
- Choice of Mistral, Hetzner, Vercel justifications
- Deployment runbook
Closed (proprietary):
- Specific prompt engineering for analysis v6 chunking
- Cache key strategy and invalidation logic
- Custom fact-check confidence scoring
- Premium features under active development (planned semi-open after market consolidation)
We're not open-source for the same reason most SaaS aren't — but we believe transparency about architecture builds trust without giving away the product.
Because the value is the pipeline (transcript extraction, chunking, domain features) not the LLM call. And because EU data sovereignty is a binding constraint for our target market. OpenAI cannot satisfy both.
Our cache layer absorbs short outages (~5-10 min) for already-analyzed videos. For new analyses, we'd return a 503 with retry guidance until Mistral comes back. We don't currently fail over to another provider — this is a deliberate tradeoff for compliance coherence (see above).
No. Mistral is called via their hosted API. For users who specifically want local LLM inference, we recommend Ollama + LM Studio + a YouTube transcript tool — it's a different product. We may offer Enterprise on-premise deployment later (not on the current roadmap).
For a 5-min video on Pro tier (mistral-medium):
- Cache hit: <500ms (Redis lookup + serialization)
- Cache miss, transcript cached: 8-15s (Mistral call)
- Cold (no cache): 25-45s (transcript extraction + Mistral)
We add maybe 500ms-2s of orchestration overhead vs raw Mistral, mostly in transcript reliability and chunking logic. Worth it.
Yes — PDF, DOCX, Markdown, XLSX. On Pro tier and above. Notion + Linear integration coming Q3 2026.
See deepsightsynthesis.com/trust (Trust Center) for subprocessor list, DPA, security whitepaper, and GDPR Art 22 disclosure.
We publish a reproducible benchmark protocol comparing Mistral against OpenAI, Anthropic and Google on our actual video-analysis task families. See BENCHMARK.md. The protocol is published first; numerical results from the first run land end of May 2026.
- Bug reports / feature requests: GitHub Issues
- Direct contact founder: maxime@deepsightsynthesis.com
- Product: deepsightsynthesis.com
Last updated: 2026-05-06. This document is versioned in github.com/Fonira/deepsight-architecture. PRs welcome on the architecture documentation itself even if the code isn't open-source.
Copyright © 2026 Maxime Leparc — Entrepreneur Individuel (SIRET 994 558 898 00015), trading as DeepSight. Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).