Handsel

Handsel helps creators and brands turn agreements into trusted payouts—structuring terms, milestones, evidence, sales reconciliation and payment release in one verifiable workflow.

Comment

The problem

Creator × brand product-line partnerships are small joint ventures run like casual conversations. Terms scatter across DMs and voice notes, no one can see who owes what next, sales are brand-reported with no shared reconciliation, and payouts land months late — or short. Creators sign on a handshake and wait 90+ days for a fraction of what they understood they'd earned. Trust evaporates the moment the handshake ends, because there is no shared, verifiable record of the promise.

What Handsel is

Handsel 信約 — named for the old Scots word for the first payment that seals a promise — is an AI-native deal-to-payout platform for creator × brand product-line partnerships. It replaces scattered messages, uncertain terms, and awkward payment follow-ups with one visible, auditable flow both sides can trust, from handshake to payout: structure the deal → sign a versioned, e-signed contract → track milestones and sales evidence on one timeline → release milestone-gated payouts via Stripe Connect.

How Gemini agents transform the workflow

Handsel runs seven specialist AI agents in production on Google Cloud, each writing an append-only event record so every judgement is replayable:

  • Deal Copilot (Google ADK · Gemini) — answers deal, contract, and milestone questions using scoped, read-only database tools.
  • Contract Intake (Vertex AI · Gemini, multimodal) — reads pasted or uploaded contracts and proposes structured settlement rules.
  • Terms Interpreter (Vertex AI · Gemini) — turns the signed contract into auditable settlement terms.
  • Sales Reconciliation (Vertex AI · Gemini + domain math) — maps messy sales exports, flags anomalies, and writes a bilingual statement; deterministic integer-only code recomputes every figure.
  • Milestone Guardian (Google ADK · Gemini · Vercel Cron) — reviews each milestone before deterministic auto-approval and freezes suspicious cases.
  • Settlement Decision (Vertex AI · Gemini + policy engine) — proposes AUTO_SETTLE / HOLD / ESCALATE_DISPUTE with confidence and citations.
  • Ops Follow-up (Google ADK · Gemini · scheduled) — scans held cases and records the next bounded action, never releasing money itself.

Google Cloud usage

Gemini is the only model path — there is no API-key fallback. General reasoning and Google ADK orchestration use Gemini 3.6 Flash; multimodal contract/image work uses Gemini 3.1 Flash Image through the global Gemini Enterprise Agent Platform endpoint. Authentication uses Application Default Credentials or Workload Identity Federation, so the deployed Next.js application does not store an AI API key. The Python bounded-agent service runs on Cloud Run (asia-east1); the model endpoint remains global, independently of the compute region.

How AI operates the business day to day

AI is not a decorative chatbot added after the workflow. New contract material enters through Contract Intake; structured terms feed the rules workflow; milestone evidence is reviewed before a deadline can auto-approve; sales rows are mapped and checked during reconciliation; a settlement agent proposes one closed disposition with citations; held cases are revisited by Ops Follow-up. The Deal Copilot gives each party a scoped view of what is due next. Every run produces safe append-only evidence — stage, bounded action, model, outcome, and timestamp — visible in the product. Scheduled agents continue monitoring milestones and held settlements rather than waiting for an operator to remember each case.

What humans do versus what AI does

Creators and brands negotiate terms, sign the contract, provide evidence, approve disputed interpretations, authorize payment mandates, and retain appeal and pause controls. Gemini reads unstructured material, finds relevant evidence, classifies risk, proposes structured rules and dispositions, and explains anomalies. Deterministic TypeScript owns amounts, authorization, idempotency, caps, contract versions, and payout eligibility. The policy engine may block an AI proposal but can never upgrade a HOLD into a payment. This separation lets AI execute meaningful operational judgement without becoming an unbounded financial actor.

Trust architecture

Trust is the product. Gemini provides judgement; deterministic code owns the money. Every payable amount is integer-only minor-unit math, ten hard policy guards sit between an agent's proposal and execution, and a model can never move funds on its own. Contracts are cryptographically hashed with RFC 3161 trusted timestamping; all state changes are append-only events; authorization is enforced at the data layer. The signed-in AI Operations console (/ai-operations) and the admin agent log (/admin/agent-log) expose agent execution evidence so a judge can replay each step from contract to settlement.

Jobs and economic opportunity

Handsel helps independent creators participate in product-line partnerships with the operating visibility normally available only to larger companies: written terms, evidence trails, understandable sales reconciliation, and predictable milestone payout status. For small brands, it reduces the coordination overhead of launching creator-led products without requiring a full legal, finance, and operations team. As adoption grows, the platform can enable paid work for creators, brand operators, production partners, accountants, and contract professionals while keeping human specialists focused on exceptions and relationships instead of repetitive record chasing.

How we built the business this way

We started with the trust boundary rather than an autonomous demo: append-only events, integer-only money, role-symmetric views, and human-confirmed contract rules. We then introduced Gemini where judgement over messy evidence creates leverage, surrounded every action with narrow tools and deterministic guards, and finally built the AI Operations console so users and judges can see that the agents really ran. That sequence — auditable foundation first, bounded autonomy second — is central to Handsel's product, not just its implementation.

What's built and live today

The full critical path — create deal → sign contract → deliver milestone → sales report → payout — is deployed to production at https://handseltrust.tech (Next.js 16, TypeScript, Vercel, Postgres, Clerk, Stripe Connect). All seven agents run in production on Google Cloud, with bilingual English / 繁體中文 support.

Honest current status

Handsel is pre-launch and pre-revenue: $0 revenue and no public users yet. We built the trust core and the AI-operations layer first, and we are now seeking our first creator × brand pairs to run real deals through the platform.

Built With

Share this project:

Updates