VeriSightAI Project Profile

Production · Live on verisightai.ai

Turn video into
verified
, publish-safe content.
VeriSight AI is not a transcription tool. Transcription is just the input. The product is trust
— source-backed claims, confidence-aware outputs, and contradiction warnings.
Python · FastAPI
Gemini 2.5 Pro/Flash
React 19 · Tailwind
MongoDB · GridFS
Stripe Payments
Whisper STT
YouTube · Vimeo · Loom
27 Iterations · 88+ Tests
See It Live
Explore the Stack

Product Identity

Not another summariser.
A trust layer
for AI content.
Every other tool in this space stops at the transcript. VeriSight starts there and goes further — verifying claims, flagging contradictions, and ensuring every published output is source-backed.
Most tools sell
VeriSight sells
Transcription / subtitles
Verification-first content intelligence
Note-taking
Source-backed claims
Generic AI summaries
Confidence-aware outputs
AI writers
Contradiction warnings & publish-safety

Pipeline

End-to-end in 8 stages.
From URL submission to publish-ready content with trust signals — fully automated, never fabricated.
01 / INPUT
Source Validation
Accepts YouTube, Vimeo, Loom URLs or direct media uploads. Rights confirmation, plan gates, quota checks run before anything else.
02 / ACQUIRE
Transcript Orchestrator
Prioritised strategy cascade: captions → Whisper audio STT → audio cleanup. Multi-adapter fallback with rotating egress proxies. Never fabricates.
03 / DERIVE
Shared Derived Representation
Single Gemini Flash call produces summary, insights, notable quotes, and claim candidates — building a reusable representation for all downstream outputs.
04 / VERIFY
Fact-Check Engine
Each claim verified against live web search (DuckDuckGo + Wikipedia). Confidence scores, verification status, and contradiction flags assigned per claim.
05 / TRUST
Trust Signal Surfacing
Badges, average confidence score, contradiction warnings. Advisory popup triggered when average confidence falls below 50%.
06 / OUTPUT
Multi-Format Delivery
5 social platforms × 3 variants, newsroom article, research brief. Lazy generation to minimise cost. Polish pass via Gemini Pro with 6 editorial styles.
07 / CACHE
Global Result Cache
30-day global dedupe: any previously-analyzed video is instantly cloned for the next user, saving AI cost and 3–10 minutes of wait time.
08 / EXPORT
Gated Exports
TXT, PDF, DOCX exports honour plan tier and watermarking rules. Per-job premium unlock ($1.50) opens all export formats for a single analysis.

Verification Contract

Hard-coded confidence rules.
No hallucination, no guessing.
Server-side rules are non-negotiable. The LLM cannot override them. Confidence values are clamped and enforced post-generation — the policy is a hard contract, not a prompt suggestion.
✓ verified
1.0
Exact confidence — green badge. Verified against live sources.
◑ partially_verified
0.10–0.90
Clamped range — amber badge. Partially supported by evidence.
? unclear
≤ 0.20
Capped — red badge. No clear supporting evidence found.
✗ contradicted
0.0
Exact zero — rose badge. Active contradicting evidence present.
◷ needs_review
≤ 0.10
Capped — slate badge. Requires human editorial judgement.

Interactive Preview

See the verification engine.
Paste a video URL and watch VeriSight classify and confidence-score extracted claims in real time.
verisightai.ai/analyze
Analyze for Trust Signals
CLAIM VERIFICATION RESULTS · 3 claims extracted
✓ VERIFIED
"Global AI investment exceeded $300 billion in 2024, with compute infrastructure accounting for the majority of spend."
confidence: 1.00 · source: Reuters, Bloomberg · 2 corroborating references
◑ PARTIAL
"Smaller open-source models now match GPT-4 on most enterprise benchmarks."
confidence: 0.61 · source: Hugging Face leaderboard · benchmark-dependent
✗ CONTRADICTED
"No major AI lab has faced regulatory fines to date."
confidence: 0.00 · contradicted by: EU AI Act enforcement record, 2025
⚠ TRUST ADVISORY · avg confidence 54% · 1 contradicted claim detected · review before publishing

Engineering

Production-grade from day one.
Every architectural decision is documented, tested, and iterable. 27 shipped iterations with zero breaking regressions.
Backend
FastAPI + Python
FastAPI 0.110
Pydantic v2
asyncio
Motor (async Mongo)
Uvicorn
BackgroundTasks
AI / LLM
Gemini 2.5 Dual-Tier
gemini-2.5-flash
gemini-2.5-pro
Whisper STT
google-genai SDK
liteLLM
token budgeting
Search & Verification
Live Fact-Check Engine
DuckDuckGo Search
Wikipedia API
yt-dlp
BeautifulSoup4
rotating egress
sha256 claim cache
Database & Storage
MongoDB + GridFS
MongoDB Atlas
GridFS attachments
compound indexes
30-day result cache
analytics_events
Auth & Payments
JWT + Stripe
PyJWT
bcrypt
magic-link email
Stripe Checkout
webhook verification
idempotent fulfil
Frontend
React 19 + shadcn/ui
React 19
React Router 7
Tailwind CSS 3
Radix UI primitives
Recharts
CRACO + Webpack
Exports & Docs
Multi-Format Output
fpdf2 (PDF)
python-docx
ReportLab
watermark engine
unicode e2e tested
Ops & Monitoring
Production Hardened
Sentry SDK
Hostinger SMTP
/health/deep endpoint
CORS lockdown
DMCA compliance
rotate_secrets.py
backend/services/transcript/orchestrator.py
# Transcript acquisition — prioritised strategy cascade, never fabricates
class
TranscriptOrchestrator
: def
__init__
(self): self
.adapters = [ PublicCaptionAdapter(), # native captions first (free, fast)
SessionBackedAdapter(), # browser-context extraction
RotatingEgressAdapter(), # rotating IPs for geo-blocked sources
AudioSTTAdapter(engine= "whisper"
), # Whisper fallback for silent caption tracks
SecondarySTTAdapter(), # secondary provider for long-form content
AlternateExtractorAdapter(), # yt-dlp last resort
] async def
acquire
(self, source: Source) -> TranscriptResult: for
adapter in
self
.adapters: if not
adapter.can_handle(source): continue
result = await
adapter.extract(source) if
result.quality_grade >= QualityGrade.ACCEPTABLE: return
result.with_provenance(adapter.label) raise
TranscriptUnavailableError( "All adapters exhausted — not fabricating"
)

Quality Assurance

88+ tests. Zero regressions across 27 iterations.
Every feature ships with a full pytest suite. Adjacent tests run on every iteration to catch regressions before they reach production.
PASSED
test_premium_unlock.py
— 9 cases: article, brief, polish, pdf, docx on premium_unlocked; 4 free-tier gate sanity checks
test_iter6_verisight.py
— 17 cases: trust advisory, contradiction warnings, confidence rule enforcement
test_orchestrator_api.py
— 10 cases: adapter cascade, quality grading, provenance tracking, fabrication guard
test_compliance_dmca.py
— URL suppression, takedown flow, re-submission blocking
test_stripe_webhook_security.py
— signature verification, bad payload, missing header, idempotency
test_video_cache_dedupe.py
— 4 cases: cache hit, fresh-URL miss, TTL-expired, incomplete prior
test_polish_article.py
— 5 cases: free block, no article 409, bad style 400, happy path, rate limit 429
test_admin_metrics.py
— 25 cases: all 11 dashboard sections, date ranges, KPI shapes, route guard
test_cost_optimization.py
— budget cap, per-user monthly limits, token estimation accuracy
test_export_unicode_e2e.py
— PDF/DOCX exports with Arabic, CJK, and emoji filenames + body content
test_support_tickets.py
— 23 cases: GridFS attachments, admin fetch, resolve + email send
test_secondary_stt.py
— Whisper long-form, engine fallback, 16-min free tier cap enforcement
test_factcheck_confidence_rules.py
— hard contract: server-side clamping, post-LLM enforcement
test_cors_required_in_prod.py
— wildcard rejection, regex fallback, production domain lockdown
test_production_config_validator.py
— boot refusal, missing keys, localhost Mongo rejection, JWT length
88 passed
in 14 test files · 0 failures · 0 warnings · adjacent regression suite green on every iteration

Plans & Pricing

Simple. Transparent. Pay-per-job available.
Every plan grants access to the full verification engine. Higher tiers unlock volume, long-form content, and API access.
Free
$0
forever
1 analysis / day · 3 / week
Watermarked copy export
Whisper ≤ 16 min
Social posts (5 platforms)
$1.00 AI budget cap
Get Started
Starter
$19
/ month
40 analyses / month
TXT export (no watermark)
Unlimited Whisper
Article + research brief
$8.00 AI budget cap
Start Starter
Pro
$49
130 analyses / month
PDF + DOCX export
1 long-form job / month (≥60 min)
Editorial polish (6 styles)
Instant cache badge
$30.00 AI budget cap
Go Pro
Agency
$149
450 analyses / month
All export formats
5 long-form jobs / month
API access + 5 seats
$100.00 AI budget cap
Contact Sales
Need just one job? $1.50 per-job credit
— unlocks all Pro features for a single analysis. No subscription required.

For Developers

Built to be extended.
Clean architecture, modular services, comprehensive admin tooling, and a health check system that tells you exactly what's wrong — not just that something is.
🔒
JWT Auth + Magic Links
bcrypt-hashed passwords, 64-char JWT secrets, email-based passwordless login. Admin bootstrap endpoint gated by BOOTSTRAP_TOKEN for secure recovery.
💳
Battle-Tested Stripe Integration
Direct stripe.Webhook.construct_event() bypassing the StripeObject shim pitfalls. Idempotent fulfillment via receipt_sent_at flags. Diagnostic UI shows per-session Stripe status.
📊
11-Section Admin Dashboard
Founder · Executive · Traffic · Funnel · Segments · Plans · Usage · Trust · Revenue · Exports · Health — with 4 date-range presets, KPI cards, and inline SVG sparklines.
⚕️
Deep Health Endpoint
/api/health/deep checks MongoDB, Stripe, SMTP, JWT length, CORS config, and more. Reports degraded state with per-component detail — not just a binary OK/fail.
🛡️
DMCA Compliance Built In
URL suppression list, takedown flow, re-submission blocking, and compliance audit trail. Fact-check cache keyed on claim SHA-256 to prevent stale verdicts.
🔄
Cost Controls at Every Layer
Two-tier LLM routing (Flash for cleanup, Pro for verdicts). Per-user monthly budget caps. Per-claim fact-check caching with 30-day TTL. Token estimation before every call.
POST /api/jobs — submit a video for verification
# Submit a YouTube URL for verification-first content extraction
import
httpx response = httpx. post
( "https://verisightai.ai/api/jobs"
, headers={ "Authorization"
: f"Bearer {token}"
}, json={ "source_url"
: "https://youtube.com/watch?v=dQw4w9WgXcQ"
, "rights_confirmed"
: True
, } ) job = response.json() # job.id → poll GET /api/jobs/{id} until status == "complete"
# job.fact_checks[0].status → "verified" | "contradicted" | "unclear"
# job.fact_checks[0].confidence → 0.0–1.0, server-enforced confidence rules
# job.posts.twitter[0] → platform-ready post, 3 variants
# job.article → full newsroom article (lazy, call POST /api/jobs/{id}/article)

Veri Sight

AI
Turn video into verified, publish-safe content.
Launch App
How It Works
Tech Stack
Pricing
For Developers
Support

Complete original text content

Production · Live on verisightai.ai
Turn video into
verified
, publish-safe content.
VeriSight AI is not a transcription tool. Transcription is just the input. The product is trust
— source-backed claims, confidence-aware outputs, and contradiction warnings.
Python · FastAPI
Gemini 2.5 Pro/Flash
React 19 · Tailwind
MongoDB · GridFS
Stripe Payments
Whisper STT
YouTube · Vimeo · Loom
27 Iterations · 88+ Tests
See It Live
Explore the Stack
Product Identity
Not another summariser.
A trust layer
for AI content.
Every other tool in this space stops at the transcript. VeriSight starts there and goes further — verifying claims, flagging contradictions, and ensuring every published output is source-backed.
Most tools sell
VeriSight sells
Transcription / subtitles
Verification-first content intelligence
Note-taking
Source-backed claims
Generic AI summaries
Confidence-aware outputs
AI writers
Contradiction warnings & publish-safety
Pipeline
End-to-end in 8 stages.
From URL submission to publish-ready content with trust signals — fully automated, never fabricated.
01 / INPUT
Source Validation
Accepts YouTube, Vimeo, Loom URLs or direct media uploads. Rights confirmation, plan gates, quota checks run before anything else.
02 / ACQUIRE
Transcript Orchestrator
Prioritised strategy cascade: captions → Whisper audio STT → audio cleanup. Multi-adapter fallback with rotating egress proxies. Never fabricates.
03 / DERIVE
Shared Derived Representation
Single Gemini Flash call produces summary, insights, notable quotes, and claim candidates — building a reusable representation for all downstream outputs.
04 / VERIFY
Fact-Check Engine
Each claim verified against live web search (DuckDuckGo + Wikipedia). Confidence scores, verification status, and contradiction flags assigned per claim.
05 / TRUST
Trust Signal Surfacing
Badges, average confidence score, contradiction warnings. Advisory popup triggered when average confidence falls below 50%.
06 / OUTPUT
Multi-Format Delivery
5 social platforms × 3 variants, newsroom article, research brief. Lazy generation to minimise cost. Polish pass via Gemini Pro with 6 editorial styles.
07 / CACHE
Global Result Cache
30-day global dedupe: any previously-analyzed video is instantly cloned for the next user, saving AI cost and 3–10 minutes of wait time.
08 / EXPORT
Gated Exports
TXT, PDF, DOCX exports honour plan tier and watermarking rules. Per-job premium unlock ($1.50) opens all export formats for a single analysis.
Verification Contract
Hard-coded confidence rules.
No hallucination, no guessing.
Server-side rules are non-negotiable. The LLM cannot override them. Confidence values are clamped and enforced post-generation — the policy is a hard contract, not a prompt suggestion.
✓ verified
1.0
Exact confidence — green badge. Verified against live sources.
◑ partially_verified
0.10–0.90
Clamped range — amber badge. Partially supported by evidence.
? unclear
≤ 0.20
Capped — red badge. No clear supporting evidence found.
✗ contradicted
0.0
Exact zero — rose badge. Active contradicting evidence present.
◷ needs_review
≤ 0.10
Capped — slate badge. Requires human editorial judgement.
Interactive Preview
See the verification engine.
Paste a video URL and watch VeriSight classify and confidence-score extracted claims in real time.
verisightai.ai/analyze
Analyze for Trust Signals
CLAIM VERIFICATION RESULTS · 3 claims extracted
✓ VERIFIED
"Global AI investment exceeded $300 billion in 2024, with compute infrastructure accounting for the majority of spend."
confidence: 1.00 · source: Reuters, Bloomberg · 2 corroborating references
◑ PARTIAL
"Smaller open-source models now match GPT-4 on most enterprise benchmarks."
confidence: 0.61 · source: Hugging Face leaderboard · benchmark-dependent
✗ CONTRADICTED
"No major AI lab has faced regulatory fines to date."
confidence: 0.00 · contradicted by: EU AI Act enforcement record, 2025
⚠ TRUST ADVISORY · avg confidence 54% · 1 contradicted claim detected · review before publishing
Engineering
Production-grade from day one.
Every architectural decision is documented, tested, and iterable. 27 shipped iterations with zero breaking regressions.
Backend
FastAPI + Python
FastAPI 0.110
Pydantic v2
asyncio
Motor (async Mongo)
Uvicorn
BackgroundTasks
AI / LLM
Gemini 2.5 Dual-Tier
gemini-2.5-flash
gemini-2.5-pro
google-genai SDK
liteLLM
token budgeting
Search & Verification
Live Fact-Check Engine
DuckDuckGo Search
Wikipedia API
yt-dlp
BeautifulSoup4
rotating egress
sha256 claim cache
Database & Storage
MongoDB + GridFS
MongoDB Atlas
GridFS attachments
compound indexes
30-day result cache
analytics_events
Auth & Payments
JWT + Stripe
PyJWT
bcrypt
magic-link email
Stripe Checkout
webhook verification
idempotent fulfil
Frontend
React 19 + shadcn/ui
React 19
React Router 7
Tailwind CSS 3
Radix UI primitives
Recharts
CRACO + Webpack
Exports & Docs
Multi-Format Output
fpdf2 (PDF)
python-docx
ReportLab
watermark engine
unicode e2e tested
Ops & Monitoring
Production Hardened
Sentry SDK
Hostinger SMTP
/health/deep endpoint
CORS lockdown
DMCA compliance
rotate_secrets.py
backend/services/transcript/orchestrator.py
# Transcript acquisition — prioritised strategy cascade, never fabricates
class
TranscriptOrchestrator
: def
__init__
(self): self
.adapters = [ PublicCaptionAdapter(), # native captions first (free, fast)
SessionBackedAdapter(), # browser-context extraction
RotatingEgressAdapter(), # rotating IPs for geo-blocked sources
AudioSTTAdapter(engine= "whisper"
), # Whisper fallback for silent caption tracks
SecondarySTTAdapter(), # secondary provider for long-form content
AlternateExtractorAdapter(), # yt-dlp last resort
] async def
acquire
(self, source: Source) -> TranscriptResult: for
adapter in
self
.adapters: if not
adapter.can_handle(source): continue
result = await
adapter.extract(source) if
result.quality_grade >= QualityGrade.ACCEPTABLE: return
result.with_provenance(adapter.label) raise
TranscriptUnavailableError( "All adapters exhausted — not fabricating"
)
Quality Assurance
88+ tests. Zero regressions across 27 iterations.
Every feature ships with a full pytest suite. Adjacent tests run on every iteration to catch regressions before they reach production.
PASSED
test_premium_unlock.py
— 9 cases: article, brief, polish, pdf, docx on premium_unlocked; 4 free-tier gate sanity checks
test_iter6_verisight.py
— 17 cases: trust advisory, contradiction warnings, confidence rule enforcement
test_orchestrator_api.py
— 10 cases: adapter cascade, quality grading, provenance tracking, fabrication guard
test_compliance_dmca.py
— URL suppression, takedown flow, re-submission blocking
test_stripe_webhook_security.py
— signature verification, bad payload, missing header, idempotency
test_video_cache_dedupe.py
— 4 cases: cache hit, fresh-URL miss, TTL-expired, incomplete prior
test_polish_article.py
— 5 cases: free block, no article 409, bad style 400, happy path, rate limit 429
test_admin_metrics.py
— 25 cases: all 11 dashboard sections, date ranges, KPI shapes, route guard
test_cost_optimization.py
— budget cap, per-user monthly limits, token estimation accuracy
test_export_unicode_e2e.py
— PDF/DOCX exports with Arabic, CJK, and emoji filenames + body content
test_support_tickets.py
— 23 cases: GridFS attachments, admin fetch, resolve + email send
test_secondary_stt.py
— Whisper long-form, engine fallback, 16-min free tier cap enforcement
test_factcheck_confidence_rules.py
— hard contract: server-side clamping, post-LLM enforcement
test_cors_required_in_prod.py
— wildcard rejection, regex fallback, production domain lockdown
test_production_config_validator.py
— boot refusal, missing keys, localhost Mongo rejection, JWT length
88 passed
in 14 test files · 0 failures · 0 warnings · adjacent regression suite green on every iteration
Plans & Pricing
Simple. Transparent. Pay-per-job available.
Every plan grants access to the full verification engine. Higher tiers unlock volume, long-form content, and API access.
Free
$0
forever
1 analysis / day · 3 / week
Watermarked copy export
Whisper ≤ 16 min
Social posts (5 platforms)
$1.00 AI budget cap
Get Started
Starter
$19
/ month
40 analyses / month
TXT export (no watermark)
Unlimited Whisper
Article + research brief
$8.00 AI budget cap
Start Starter
Pro
$49
130 analyses / month
PDF + DOCX export
1 long-form job / month (≥60 min)
Editorial polish (6 styles)
Instant cache badge
$30.00 AI budget cap
Go Pro
Agency
$149
450 analyses / month
All export formats
5 long-form jobs / month
API access + 5 seats
$100.00 AI budget cap
Contact Sales
Need just one job? $1.50 per-job credit
— unlocks all Pro features for a single analysis. No subscription required.
For Developers
Built to be extended.
Clean architecture, modular services, comprehensive admin tooling, and a health check system that tells you exactly what's wrong — not just that something is.
🔒
JWT Auth + Magic Links
bcrypt-hashed passwords, 64-char JWT secrets, email-based passwordless login. Admin bootstrap endpoint gated by BOOTSTRAP_TOKEN for secure recovery.
💳
Battle-Tested Stripe Integration
Direct stripe.Webhook.construct_event() bypassing the StripeObject shim pitfalls. Idempotent fulfillment via receipt_sent_at flags. Diagnostic UI shows per-session Stripe status.
📊
11-Section Admin Dashboard
Founder · Executive · Traffic · Funnel · Segments · Plans · Usage · Trust · Revenue · Exports · Health — with 4 date-range presets, KPI cards, and inline SVG sparklines.
⚕️
Deep Health Endpoint
/api/health/deep checks MongoDB, Stripe, SMTP, JWT length, CORS config, and more. Reports degraded state with per-component detail — not just a binary OK/fail.
🛡️
DMCA Compliance Built In
URL suppression list, takedown flow, re-submission blocking, and compliance audit trail. Fact-check cache keyed on claim SHA-256 to prevent stale verdicts.
🔄
Cost Controls at Every Layer
Two-tier LLM routing (Flash for cleanup, Pro for verdicts). Per-user monthly budget caps. Per-claim fact-check caching with 30-day TTL. Token estimation before every call.
POST /api/jobs — submit a video for verification
# Submit a YouTube URL for verification-first content extraction
import
httpx response = httpx. post
( "https://verisightai.ai/api/jobs"
, headers={ "Authorization"
: f"Bearer {token}"
}, json={ "source_url"
: "https://youtube.com/watch?v=dQw4w9WgXcQ"
, "rights_confirmed"
: True
, } ) job = response.json() # job.id → poll GET /api/jobs/{id} until status == "complete"
# job.fact_checks[0].status → "verified" | "contradicted" | "unclear"
# job.fact_checks[0].confidence → 0.0–1.0, server-enforced confidence rules
# job.posts.twitter[0] → platform-ready post, 3 variants
# job.article → full newsroom article (lazy, call POST /api/jobs/{id}/article)
Veri Sight
AI
Turn video into verified, publish-safe content.
Launch App
How It Works
Tech Stack
Pricing
Support