ponti

The bridge between brands and creators. Fair pay, Vetted talent, Zero admin.

Comment

Inspiration

In Nairobi, a creator ships a video, posts it, and then waits. The brand "will review it." Weeks pass. Sometimes the money comes, as an awkward M-Pesa request to a stranger. Sometimes it just doesn't. Meanwhile the brand is drowning: manually watching dozens of submissions, never sure the creator they picked actually followed the brief. The talent is here. The plumbing is broken.

So we asked: what if an AI agent sat in the one seat that's always been human, slow, and unfair: the seat that decides whether a creator's work is good enough to get paid?

What it does

Ponti is income infrastructure for the informal creator economy, and an AI agent is what makes it work.

Here's the loop. A brand funds a campaign, and the money goes into escrow before a creator lifts a finger. The creator accepts, shoots, and uploads. Then Gemini, on Vertex AI, actually watches the footage. plus the caption. It reads the brief as a checklist, checks each requirement against what it literally sees on screen (it even reads the hashtags overlaid inside the video), and returns a verdict (APPROVED, REVISION, or FLAG) with a confidence score and cited evidence for every line. Approve, and the creator gets paid their full rate. No chasing. No ghosting.

The agent is the product, not decoration on a dashboard. The most expensive, dispute-prone judgment in influencer marketing ("did they deliver what we paid for?") is the one we handed to the AI. And it never gets the last word over money: below 78% confidence, the decision routes to a human.

Around that agent sits the rest of the machine: escrow that holds the brand's budget safely, per-platform verified creator profiles (brands match on real reach, not bought followers), promo-code commission tracking, and a one-page readout the brand can forward upward. The money and the proof of work live in one place. WhatsApp threads don't.

For a continent where creators are routinely underpaid and paid late, Ponti turns "get paid eventually, maybe" into "get paid on approval, verifiably."

How we built it

Stack: Next.js 14 (App Router) · TypeScript · Neon Postgres (serverless) · Vertex AI (Gemini 2.5 Flash) · Cloud Run · Cloud Storage · Resend · Firebase Auth.

Why Gemini: the assessment agent must watch a video and read a text brief at once. Native multimodal is the right tool for the single most expensive judgment in the pipeline. The AI sits at that decision, not on a dashboard.

The state machine: every campaign runs a defined lifecycle (DRAFT → PENDING_REVIEW → FUNDED → CONTRACT_SIGNED → SUBMITTED → UNDER_REVIEW → APPROVED → POSTED → PAID) with failure branches at every node: a missed deadline auto-expires and refunds the brand; a low-confidence AI call routes to a human instead of auto-deciding over money.

Build wins: direct-to-GCS signed uploads (videos never touch our server); an ACID escrow ledger in whole shillings (no floating-point drift); graceful AI degradation (Vertex → Groq → transparent heuristic, so a submission is never lost); fire-and-forget email (a mail outage never stalls a campaign).

Challenges we ran into

  • Verification without invasive asks: we started with a paste-a-code-in-bio flow; creators found it clunky, so we moved to OAuth account-linking (TikTok Login Kit, live) that verifies ownership at the source.
  • Escrow + regulation: holding brand funds in Kenya may count as payment intermediation; we run pilots under written agreements with founding brands while we clarify the path. We'd rather name the open question than pretend it isn't one.
  • AI confidence calibration: tuning the human-review fallback below 78% confidence. An AI that auto-decides over real money with no override is a liability, not a feature, so we built the uncertainty path first.
  • Pricing transparency: creators asked how they'd know a campaign pays fairly, so we show them the campaign rate vs. their own per-platform rate card before they apply. Nobody discovers the number after saying yes.
  • Registration on the critical path: live payments and production API access both depend on a registered Kenyan business, and the BRS approval took longer than we planned. It held up the money rail, so we ran early campaigns under written agreements with founding brands while we waited on it.
  • Money that doesn't leak on the way through: every mobile-money and bank rail charges a fee, so picking accounts and settlement paths that don't quietly erode a creator's rate, while staying inside the government's shifting payments rules, is a constant balancing act.

Accomplishments that we're proud of

  • We got registered. Ponti is now a registered business in Kenya (eCitizen BRS). That was a hard dependency for live payments and production integrations, and clearing it turns the revenue path from "planned someday" into "unblocked now."
  • A live AI agent that gates real money. On real creator submissions it approved a compliant cupcake review at 100% confidence, correctly reading the brand's hashtag rendered inside the video, and rejected a non-compliant one at 0%. Right both times.
  • Human-in-the-loop, built first. The below-78%-confidence override path shipped before the auto-approve path. The safety rail came before the feature.
  • Money that balances exactly. An ACID escrow ledger in whole shillings, direct-to-GCS uploads, and graceful AI degradation so a submission is never lost.
  • Ownership verified at the source via TikTok Login Kit OAuth. Not a screenshot, not an honor system.
  • We told the truth about what's live. A clear real-vs-roadmap line a judge can verify against the repo (see below).

What we learned

  • Put the AI at the decision, not on the dashboard. Native multimodal is the right tool for the single most expensive judgment in the pipeline; wrapping a model around a UI isn't the same as trusting it with the call.
  • Friction kills verification. Paste-a-code-in-bio felt invasive and clunky; OAuth at the source is both easier and more trustworthy.
  • Name the open question. Escrow may be payment intermediation in Kenya. Saying so plainly and running written-agreement pilots beats pretending the risk isn't there.
  • Honesty compounds. A transparent heuristic fallback and an explicit real-vs-roadmap line earn more trust, from creators and judges alike, than an over-claim ever could.

What's real vs. roadmap (so a judge can check): the submission-assessment agent (Gemini watching the video) gates payment and runs in production today. The creator Trust Score and creator↔campaign matching are rule-based/heuristic today. AI-assisted creator vetting is on the roadmap; we do not claim it as live.

What's next for ponti

  • Live payments via IntaSend (M-Pesa + bank) now that Ponti is a registered business in Kenya: real escrow collect and payout, with KYC at the point of payout.
  • AI-assisted creator vetting to complement the human/heuristic Trust Score at onboarding.
  • Pay-per-view (CPM) campaigns and an accept/reject join gate, so brands with a fixed budget can cap and choose who joins.
  • Automated post-verification (confirming the live post is real and matches the verified handle), plus more platforms through Login-Kit-style OAuth.
  • Grow the founding-brand pilots from the first campaigns into a repeatable, self-serve flow.

Built With

Share this project:

Updates