Full-Stack Developer for AI Agentic Finance App – Supabase
FreelanceJobs
- Canada
- Canada
À propos
The frontend is already done.
I have complete Figma designs and the frontend has already been converted to React Native components. You will not be building UI from scratch. Your job is to integrate these components with the backend services and build out the backend architecture that powers the agentic AI. This is a backend-heavy role with some frontend wiring.
Critical Architectural Requirement: Deterministic Data + Agentic AI Layer
The app must maintain a strict separation between deterministic financial data and the agentic AI layer. This is non-negotiable:
Deterministic data (transactions, balances, budgets, bills) is the source of truth. It must be immutable, auditable, and mathematically verifiable. Every change must be logged with a clear audit trail.
The agentic AI can read this data to answer questions and propose actions, but it can never directly modify deterministic data. All write operations must go through a confirmation flow where the user explicitly approves, and the change is recorded as a separate auditable event.
The AI must cite the specific transactions that support every answer. If it cannot find the answer in deterministic data, it must say so – never invent.
This separation ensures users can trust the numbers absolutely, while still benefiting from AI assistance. It also enables future products (white‑label for banks, financial integrity layers) without a rewrite.
Technical Stack
Frontend: React Native (existing components)
Deterministic Backend: Supabase (Postgres, RLS, Functions)
Live Banking: Plaid
Agentic AI: OpenAI ChatGPT API + Function Calling
Subscriptions: RevenueCat
Attribution: Meta Pixel SDK
Hosting: Vercel/Netlify + Supabase
Scope of Work
Deterministic Backend & Architecture
Set up Supabase with a multi‑tenant data model (all tables scoped by tenant_id).
Implement Row Level Security (RLS) policies for complete tenant isolation.
Design deterministic database schema for:
Users and authentication
Linked accounts (Plaid)
Transactions (with pending → posted reconciliation, duplicate suppression, immutable history)
Categories (system + user-defined)
Budgets (with versioning so changes are traceable)
Bills and recurring items
Sync logs and change history (every data change is recorded)
Ensure immutable storage where appropriate – past transactions should never be overwritten, only corrected via audit trails.
Implement row-level hashing or snapshots to detect retroactive changes (future-proofing for audit readiness).
Integrations
Plaid: Connect live bank accounts, sync transactions, handle webhooks, manage connection health (Connected / Attention Required). Ensure deterministic reconciliation – no duplicates, no missed transactions.
Agentic AI (OpenAI): Build secure Edge Functions that:
Accept user questions and action requests
Fetch user's deterministic transaction data from database
Use function calling to enable the AI to propose actions (categorize transaction, set budget, add bill)
Construct guardrailed prompts that enforce citation and prevent hallucination
Return answers with citations linked to specific transaction IDs
Never write directly to the database, instead, return structured action proposals that the frontend presents for user confirmation
Log all AI interactions (both reads and action proposals) for audit
RevenueCat: Integrate subscriptions, handle webhooks, sync state with Supabase.
Meta Pixel SDK: Implement attribution events on frontend.
Frontend Integration
Connect the existing React Native components to Supabase (auth, data fetching, real-time updates).
Implement Plaid Link flow (open, handle success/error).
Add RevenueCat SDK for paywall and subscription management.
Add Meta Pixel SDK and track key events.
Build confirmation modals for AI‑proposed actions – user must tap "Confirm" before any write operation is sent to Supabase.
Backend Functions
Build Edge Functions for:
Agentic AI chat (with function calling, read‑only access to deterministic data)
Plaid webhooks (idempotent, deterministic)
RevenueCat webhooks
Complex aggregations (cash flow, spending summaries) that read from deterministic data
Ensure functions are secure, logged, and performant.
Testing & Documentation
Unit/integration tests for critical logic:
Sync reconciliation (deterministic matching)
AI guardrails (ensuring citations and no hallucination)
Action confirmation flow (AI proposes → user confirms → database write)
Subscription updates
Clear documentation: database schema, functions, environment variables, deployment steps, RLS policies, audit trail design.
Deployment Assistance
Configure production environment (Supabase, frontend hosting).
Ensure security best practices and scalability.
Smooth handoff with documentation.
What I Already Have (You Don't Build)
Figma designs for all screens (Dashboard, Transactions, Budgets, Cash Flow, AI chat, confirmation modals, etc.)
React Native components converted from Figma – production‑ready and waiting to be connected
Detailed PRD covering user flows, acceptance criteria, agentic AI persona rules, deterministic data requirements, action confirmation patterns, analytics events
Clear technical requirements (this document)
Must-Haves for Applicants
Proven experience with Supabase (or PostgreSQL + RLS) and multi‑tenant application design.
Deep understanding of deterministic data patterns – immutable audit trails, idempotency, row-level versioning, and reconciliation logic.
Experience integrating Plaid (or similar financial APIs) and handling webhooks with idempotency guarantees.
Experience with OpenAI API including function calling to enable agentic behavior.
Deep understanding of guardrails to prevent hallucination and ensure AI never invents financial data.
Familiarity with RevenueCat and mobile subscription flows.
Strong frontend skills in React Native to connect existing components.
Strong security practices (environment variables, least privilege, audit logging).
Clean, documented, testable code.
Strong English communication (daily collaboration via Slack/Upwork).
Screening Questions (Please Answer in Your Proposal)
To help me identify the best fit, please answer these specific questions:
Deterministic Data Design: You're building a transaction table that must maintain an immutable history. A transaction from 2 weeks ago was incorrect and needs to be corrected. How do you handle this without deleting or overwriting the original record? Show a sample database schema and explain the audit trail.
Multi‑tenancy & RLS: You have a Supabase app with multiple tenants. A new tenant is created and needs to see only its own data. Walk me through the exact steps you'd take – from database design to RLS policy creation – to ensure complete isolation. Include a sample policy.
AI Hallucination Prevention + Determinism: You're building an agentic AI that answers questions about a user's deterministic transaction data. Describe your approach to guarantee the AI never invents a transaction or balance. How do you structure the prompt? What backend validation do you add before showing the answer to the user to ensure every number can be traced back to a database record?
Function Calling with Deterministic Safeguards: The user says: "Move my Starbucks transactions from yesterday to the 'Coffee' category." Walk me through how you'd implement this using OpenAI function calling. Include the function definition, how you validate the AI's proposed action against deterministic data, and how you ensure the actual database write only happens after user confirmation.
Plaid Webhook Reliability (Idempotency): Plaid sends webhooks for transaction updates. How do you ensure that no transaction is missed or double‑counted, even if webhooks arrive out of order or the same webhook is sent multiple times? Outline your idempotency strategy and how you maintain deterministic sync state.
RevenueCat + Supabase Sync: A user subscribes via RevenueCat. How do you keep the user's entitlements in sync with Supabase? Describe the webhook handling flow and how you'd handle edge cases like a lapsed subscription or a refund.
React Native Integration: You have a pre‑built React Native component for the transaction list. Walk me through how you'd fetch data from Supabase, handle loading states, and implement real‑time updates when new deterministic transactions arrive via Plaid webhook.
Audit Trail for Agentic Actions: The AI proposes an action, the user confirms, and the action is executed. Design a database table schema for tracking this entire flow – from proposal to confirmation to execution – ensuring a complete audit trail that ties back to the deterministic data that was changed.
Handling Retroactive Changes: A processor (or bank) sends a corrected transaction that changes a balance from 2 months ago. How does your system handle this while maintaining deterministic truth? How do you notify the user and ensure the AI doesn't give contradictory answers based on stale data?
Key Deliverables
Fully functional backend with Supabase, Plaid, OpenAI (with function calling), RevenueCat, Meta Pixel integrated.
Deterministic data foundation with immutable audit trails, idempotent webhook handling, and row-level versioning.
Multi‑tenant architecture with RLS in place (even though initial launch is single‑tenant, the structure supports adding tenants later).
Agentic AI that:
Answers questions with transaction citations (every number traceable to deterministic data)
Proposes actions (categorize, budget, bill management) via function calling
Requires user confirmation before any write
Never hallucinates financial data
Webhook handlers for Plaid and RevenueCat (idempotent, deterministic).
Frontend fully connected to backend services with confirmation modals.
Comprehensive audit logging for all AI‑proposed and user‑confirmed actions.
Documentation covering schema, functions, deployment, and the deterministic data architecture.
I'm looking for a long‑term partner. If this goes well, there will be ongoing work: new features, white‑label versions, scaling.
How to Apply
Please include in your proposal:
A brief intro and relevant experience (specifically mentioning Supabase, Plaid, OpenAI function calling, React Native, multi‑tenancy, and deterministic data patterns).
Links to similar projects / portfolio.
Answers to the screening questions above.
Your proposed milestone breakdown and timeline.
Any questions or suggestions about the stack or scope.
I'll share the PRD and designs with shortlisted candidates under NDA.
I'm excited to find a true partner who understands the importance of deterministic financial data and can build an agentic AI on top of it. Let's build something great
Contract duration of 1 to 3 months.
Mandatory skills: React, API Integration, Artificial Intelligence,
Compétences linguistiques
- English
Avis aux utilisateurs
Cette offre provient d’une plateforme partenaire de TieTalent. Cliquez sur « Postuler maintenant » pour soumettre votre candidature directement sur leur site.