KanDelta

Opportunity intelligence for the video internet

"Find the delta before the market does."

KanDelta finds where audience attention and creator supply diverge, using real video intelligence from Oriane, built and deployed on Replit.

The name. KAN = Team Kanban. Delta (Δ) = the divergence between creator supply and audience response.

Oriane understands the video internet. KanDelta identifies the opportunity hidden inside it.

KanDelta by Team Kanban · https://github.com/4waiz/kandelta

Problem

Brands, agencies and creators decide what to make by scrolling what's already trending. By the time a format is visible enough to copy, it's crowded: everyone ships the same POV, the same talking head, the same unboxing. Social analytics tools read captions and hashtags, so they miss most of what actually happens inside the video.

Would someone pay for this? A brand planning a creator campaign spends thousands per video. Knowing which creative angle audiences reward, before the crowd arrives, and having the evidence, the recipe, the creators and the brief in one place, replaces days of manual research.

Insight

When audience response is unusually high but creator supply is still low, there is a creative opportunity. KanDelta measures that mismatch with real video intelligence, not guesses.

What it found (real Oriane data, last 3 months)

Search "Running shoes UAE" →

• Oriane measured 48,197 running videos (Instagram + TikTok).
• Desert heat (runners outdoors under hot desert sun) appears in only 119 videos, 0.25% of supply, yet earns 2.1× the engagement per view of the average running video.
• Found by watching, not reading: only 12 of those 119 videos mention heat, hot weather or humidity in their caption or speech. A caption or transcript search would miss ~90% of this opportunity. Oriane's vision sees it in the frames.
• Brands already pay for this look: several of the most-watched examples are disclosed ads or boosted posts (e.g. partner posts tagging @altrarunning, @zalando). Organic creators barely make it.
• No brand owns the space: the six biggest running brands collect 2,068 mentions across running videos, but only 33 of those mentions are in desert-heat videos.
• Crowd gap: 49% of style-matched running videos are first-person POV; desert heat is 9% of them and earns 2× the engagement per view.

How it works

Oriane (vision · spoken words · metadata · comments)
  → KanDelta measurement (supply, reach, engagement, momentum)
  → Opportunity map → Evidence → Crowd Gap → Creative DNA
  → Audience Lab → Creator Fit → Campaign Brief → Activate (TikTok · Reels · Shorts)

1. Discover. Describe a market. KanDelta measures ~28 creative angles: 22 detected in what's said (exact phrases in the transcript or caption) and 6 detected in what's shown (Oriane visual similarity ≥ 0.8 against text prompts such as "a person outdoors under hot desert sun").
2. Delta Map. Supply (x) vs audience response (y). Top-left = the opportunity delta.
3. Evidence. Real videos, ranked by visual match, with sponsored / likely-boosted posts separated and a brand-safety screen on what's displayed.
4. Crowd Gap. What most creators default to vs what audiences reward (visual style, sound, collabs, tagging, language). A gap is only reported when the data shows one.
5. Who's already there + Hidden conversation. Brand coverage of the gap, and what's spoken about brands in videos that never caption or tag them (Oriane Shadow Reach concept, via transcript-include + caption-exclude + mention-exclude filters).
6. Creative DNA. Measured on the evidence: talk-first openings (transcript timestamps), duration, voice density, original vs licensed audio, CTA patterns, real hooks (first 3 seconds, transcribed by Oriane), audience comments.
7. Audience Lab. Pressure-test the brief with six simulated personas before paying for production.
8. Creator Fit. Creators ranked by proven execution: the angle video's views vs the creator's own median (fetched per creator), reach, engagement, on-topic share and freshness. Not follower count.
9. Activate. A concrete brief plus TikTok / Reels / Shorts adaptations, and an Evidence Vault swipe file.

Why Oriane

Oriane is the video intelligence layer. KanDelta uses it deeply:

API notes (non-secret): docs/ORIANE_API_NOTES.md.

Why Replit

KanDelta runs and deploys on Replit: Next.js on Node 20, ORIANE_API_KEY in Replit Secrets, autoscale deployment (.replit). The Oriane cache of real responses ships with the repo, so the demo stays reliable even if the API is slow.

Architecture

browser ──▶ Next.js UI (React 19, Tailwind 4)
              │
              ▼
        /api/analyze · /api/opportunity · /api/landscape   (server routes)
              │
              ▼
        lib/whitespace/*  (scoring, crowd gaps, DNA, creators, brief, audience lab)
              │
              ▼
        lib/oriane/client.ts  (server-only; memory → disk cache → Oriane API; stale fallback)
              │
              ▼
        connect.oriane.xyz  (Bearer ORIANE_API_KEY, never sent to the browser)

Opportunity model

All inputs are Oriane population statistics. The 3 most-viewed videos are excluded from every rate (views, followers, interactions), so one viral hit or boosted ad cannot manufacture an opportunity.

Reach index         = (views ÷ followers)_angle ÷ (views ÷ followers)_market
Engagement index    = (interactions ÷ views)_angle ÷ (interactions ÷ views)_market
Response            = √(reach index × engagement index)
Performance         = clamp(0.5 + log₂(response), 0, 1)
Scarcity            = 1 ÷ (1 + supply share ÷ 1%)
Confidence          = size factor (log₁₀ n ÷ log₁₀ 300) × concentration factor (top-3 share of views)
Opportunity Score   = 100 × √(Performance × Scarcity) × Confidence
Stage               = EARLY < 0.5% ≤ EMERGING < 1.5% ≤ CROWDED < 4% ≤ SATURATED   (current, not a forecast)
Momentum            = angle's last-30-day share of its 3-month supply ÷ the market's

Every screen has "How is this calculated?".

Crowd Gap

For each dimension KanDelta compares the most common choice (creator consensus) with the best-responding choice (audience response). A gap is reported only when the best-responding choice beats the consensus by ≥ 1.25×. Otherwise it says "No gap here".

Audience Lab (honest by design)

Simulated audience panel: directional feedback, not real market research. Six rule-based personas (serious runner, casual student, potential buyer, content creator, heavy short-form viewer, UAE resident) react to measurable features of the draft (when the stakes are stated, jargon, proof beats, price context, CTA style, local and Arabic cues), weighed against norms Oriane measured on the evidence. Every reaction lists why it fired; recommendations cite their evidence. It never produces market statistics such as "87% would buy".

Running locally

npm install
cp .env.example .env.local   # then set ORIANE_API_KEY
npm run dev                  # http://localhost:3000

• npm run demo:fixtures serves only saved real Oriane responses (no API calls, no credits).
• npm run explore -- scripts/markets/running.json is the exploration harness used to validate the idea.
• npm run lint, npm run typecheck, npm run build.

Environment variables

Hackathon

Oriane x Replit: Build for the Video Economy, Dubai, 27 September 2026. Built today.

Limitations

• KanDelta does not predict virality. It identifies observed mismatches between content supply and audience response.
• Angle supply depends on how an angle is detected: phrase matching misses videos that never say it, and vision prompts are approximations (threshold 0.8).
• Oriane aggregates are sums, so large accounts weigh more; we trim the top 3 videos and show confidence.
• Market definition is caption-based ("running" also catches "running late"); evidence is screened for relevance and brand safety, but population numbers include some noise.
• Location slices (e.g. UAE-only) are often too small; KanDelta says so and falls back to the global market.
• Audience Lab is simulated. YouTube Shorts adaptations are not grounded in Oriane data (Oriane indexes Instagram + TikTok).
• Oriane API credit costs are not publicly documented; results are cached aggressively.

Naming note

The product was renamed from its working title during the hackathon. Internal module paths (src/lib/whitespace), environment variables (WHITESPACE_*) and cache/storage keys intentionally keep the old name so the tested build and the cached real Oriane responses stay valid. They will be renamed after the hackathon.

Future work

• Live Signal: continuously monitor fresh Oriane data and alert when a new crowd gap opens (not built; we don't fake live monitoring).
• Demand-gap mining from comments at scale (recurring questions → supply check).
• Image-asset vision prompts (logo / product shots) for visual Shadow Reach.
• Expose KanDelta as a tool for AI agents (Oriane already ships MCP for AIs).