-
-
-
-
-
-
-
-
-
-
Hero
Inspiration
Freelancers do more than deliver work. They translate vague briefs, negotiate scope, track deadlines, collect feedback, and chase payments.
We built Tranche to connect these responsibilities in one workspace. Our goal was to help freelancers and agencies move from a client brief to verified, paid milestones while keeping people in control of important decisions.
What it does
Tranche turns client briefs into AI-generated project plans with milestones, tasks, budgets, and delivery requirements.
Freelancers can review the plan, schedule dependencies in an interactive Gantt chart, and share it through a client portal. A negotiation room makes agent activity and requests for human review visible.
When work is submitted, Tranche checks delivery evidence against milestone requirements. Payment actions follow verification and approval rules, with PayPal handling milestone authorization and capture.
A cash-flow dashboard shows held funds, released payments, and forecasts. An audit trail records project decisions and agent activity.
How we built it
We built Tranche as a TypeScript monorepo using pnpm and Turborepo.
The web application uses Next.js, with PostgreSQL and Drizzle for persistent data. Gemini powers AI planning and assessment through the Vercel AI SDK, while Render Workflows supports background execution.
Bryntum Gantt provides interactive scheduling, and AG Grid Studio powers the cash-flow dashboard.
We separated AI recommendations from financial execution. Agents submit typed proposals instead of calling money-moving PayPal endpoints directly. A deterministic policy engine evaluates those proposals, and an outbox handles PayPal writes with idempotency keys.
The hackathon payment workflow runs in PayPal Sandbox.
Challenges we ran into
The biggest challenge was connecting probabilistic AI output to a payment workflow that requires predictable behavior. We addressed this with structured outputs, machine verification, explicit approval gates, and a strict separation between proposals and execution.
Keeping schedules, delivery states, client decisions, and payment status consistent across asynchronous workflows was another challenge. We used explicit state transitions and traceable events to make those changes understandable.
We also encountered model quota limits and misleading error messages. Improving model configuration and error handling helped the application explain failures more clearly.
Accomplishments that we're proud of
We completed a real USD 1.00 PayPal Sandbox milestone payment after delivery verification and client release.
We demonstrated a failed delivery check followed by a corrected submission, showing that verification can block progress when requirements are not met.
We brought planning, scheduling, client review, delivery evidence, payments, and cash-flow visibility into one working application.
Most importantly, we kept human oversight central to the workflow. AI can assist with decisions, but it cannot override failed machine checks or directly move money.
What we learned
Useful AI automation needs clear boundaries, visible evidence, and a reliable path to human review.
We learned that payment workflows require careful handling of retries, duplicate events, and partial failures. A successful API response is only one part of a trustworthy end-to-end experience.
We also learned that transparent status messages and readable audit trails matter as much as the automation itself. Users need to understand what happened, what is waiting, and what requires their attention.
What's next for Tranche
We want to test Tranche with freelancers and small agencies to improve briefing, milestone planning, and client approvals.
Our next priorities include broader delivery verification, clearer negotiation summaries, better notification visibility, and stronger recovery from interrupted workflows.
We also plan to expand automated testing across rejected deliveries, schedule changes, and payment edge cases before pursuing production payment readiness.
Our goal is to make freelance projects easier to manage, easier to verify, and easier to get paid for.
Built With
- ag-grid-studio
- better-auth
- bryntum-gantt
- docker
- drizzle-orm
- github-api
- google-gemini
- google-pagespeed-insights
- next.js
- paypal
- paypal-javascript-sdk
- playwright
- pnpm
- postgresql
- react
- tailwind-css
- turborepo
- typescript
- vercel-ai-sdk
- zod
