The AI manager that turns messy brand email into
structured sponsorship decisions.
Real Deal is a Gmail-connected AI sponsorship manager for creators who receive brand outreach before they can justify a full-time manager. It reads conversations, verifies risk, extracts terms, saves the agent output, and turns scattered inbound mail into safe next actions the creator can approve.
Exact Gemini models used
gemini/gemini-2.5-flash-litegemini/gemini-2.5-flashgemini/gemini-2.5-pro9
named agent roles
Gmail
thread analysis
Gemini
models only
Supabase
saved outputs
The problem
The inbox is where creator businesses leak money.
Sponsorship work looks simple from the outside: a brand emails, a creator replies, and a deal happens. The real workflow is messier. A creator has to judge legitimacy, ask for budget, decode deliverables, protect usage rights, remember past rates, watch for scams, and respond in a tone that keeps the relationship warm.
That is repetitive business judgment, not casual email. It is also the exact work that creators do when they are already busy making content. The result is decision debt: good deals wait too long, weak deals take too much attention, and risky emails slip through because every thread looks urgent.
Every opportunity arrives incomplete
A brand may mention a product but not the budget. An agency may ask for usage rights without saying how long. A gifted-only offer can look friendly while still wasting the creator's time.
Creators undercharge for hidden terms
Paid usage, whitelisting, rush deadlines, exclusivity, revisions, and platform-specific deliverables change the value of a deal. Most inbound emails do not make those costs obvious.
Scams imitate legitimate sponsorships
Lookalike domains, strange links, pressure tactics, unknown senders, and vague payment language create risk. Smaller creators often have to make those calls without a manager.
A brand deal email is rarely just one email. It is a negotiation thread with money, risk, memory, and timing hidden inside it.
Real Deal design principle
The system
Nine agents, one guarded business workflow.
The backend is built around CrewAI-style specialist roles and production flows. Agents do not all try to answer the same prompt. Each one owns a slice of the business decision: intake, risk, extraction, fit, rate intelligence, negotiation, drafting, memory, and final approval.
Gemini models power the reasoning. Faster Gemini models handle low-cost classification and extraction, while deeper Gemini models are reserved for writing and high-stakes strategy where quality matters more than speed.
Inbox Intake Agent
normalizationReads Gmail threads, groups messages into conversations, extracts sender data, timestamps, snippets, headers, and thread context before any business decision is made.
Spam and Scam Detection Agent
riskChecks impersonation signals, suspicious domains, risky links, strange payment requests, urgency pressure, and whether the sender identity matches the claimed company.
Deal Extraction Agent
termsPulls out brand, product, ask, deliverables, budget, timeline, usage rights, exclusivity, category, and missing terms as structured data.
Brand Fit Agent
alignmentCompares the offer with available creator preferences, category history, known brand context, and persisted collaboration signals when they exist.
Rate Intelligence Agent
pricingLooks at learned rate rules and accepted deal evidence when present, then flags missing price context, usage rights, and exclusivity before a reply is drafted.
Negotiation Strategy Agent
next moveDecides whether to ask for info, counter, reject, verify, wait, escalate, or draft a reply based on the conversation state and deal quality.
Reply Draft Agent
languageCreates reviewable reply drafts and runs them through placeholder cleanup so bracketed names, fake contacts, and vague signoffs do not reach approval.
Memory Agent
learningWrites analysis memory events for every run, then updates brand, category, tone, and rate memory only when the thread contains real sponsorship evidence.
Compliance Gatekeeper Agent
controlBlocks unsafe automation for money terms, legal uncertainty, scam risk, low confidence, usage rights, exclusivity, or anything that needs the creator's explicit approval.
Conversation analysis
The thread is the source of truth.
Real Deal does not treat every new Gmail message as a separate deal. It loads the conversation first. If the creator already replied, if the brand already accepted terms, if the deal is finished, or if the latest message is just a notification, the agent should understand that before it recommends anything.
This matters because sponsorships are negotiated across time. One message may contain the budget, another may contain the deadline, another may reveal usage rights, and the final reply may show the creator's accepted price. A useful manager remembers the whole exchange.
Agent run trace
CrewAI workflow · Gemini models · streamed with safe summaries
Conversation is loaded
The backend reads the whole Gmail thread first, not just the newest email, so replies, accepted terms, and already-finished deals are understood before drafting.
Deal terms are extracted
Budget, ask, deliverables, timeline, usage rights, exclusivity, payment terms, category, and sender identity are converted into structured outputs.
Risk and fit are scored together
The system checks whether the sender is safe, whether the product fits, and whether the email is a real sponsorship, affiliate pitch, gifted offer, or non-deal message.
Memory changes the recommendation
If previous conversations have produced learned rates, creator preferences, or brand history, that stored context can shape the next move.
A safe next action is produced
The final output explains what to do next: ask for budget, counter with a price, wait for the brand, reject politely, verify identity, or archive.
Creator approval remains the gate
Drafts can be edited, approved, or rejected. High-stakes actions are never silently sent just because an agent produced a confident result.
Memory
Memory is what makes this more than email automation.
A generic email assistant can summarize a message. Real Deal is designed to keep a persistent record of how sponsorship conversations were handled. It stores useful signals in Supabase: agent runs, Gmail thread records, deal assessments, draft replies, analysis memory events, accepted rates when present, categories, and creator tone signals from resolved deal threads.
The important boundary is that memory has to be honest. A Google security alert, a school notice, or a billing receipt can be saved as an audit event, but it should not become sponsorship memory. The backend only updates sponsorship deal memory when the conversation state and extracted terms provide enough evidence.
01
Assessment memory
Each analysis can persist the score, confidence, budget, category, missing terms, next action, and conversation state as an audit trail.
02
Deal memory
Only accepted, completed, or clearly rejected sponsorship conversations update brand and creator memory. Non-deal emails stay out of pricing intelligence.
03
Rate memory
When a completed or accepted thread includes a usable budget, the backend can save learned rate rules tied to the deliverable and source assessment.
04
Tone memory
When the creator has replied inside a resolved deal thread, the backend can infer a small tone profile from those creator-written messages.
Product surface
Every page has one operational job.
Dashboard
A compact operating view for inbox health, analyzed vs waiting conversations, deal readiness, and the next batch of messages that need agent analysis.
Inbox
A live Gmail workspace with filters, real message ordering, per-email analysis status, selected-thread analysis, and an agent suggestion panel.
Drafts
A review queue for generated replies, including edit and approval flows plus a mobile swipe mode for approving or revising one draft at a time.
Deals
A sponsorship-focused pipeline that shows qualified opportunities, stages, missing terms, risk holds, and persisted assessment history.
Profile
A creator control center for Gmail access, learned preferences, rate memory, feature readiness, automation gates, and account identity.
Agent logs
A live run page streams CrewAI and Gemini events as readable summaries so judges and creators can see the workflow without exposing hidden chain-of-thought.
Safety and control
The agent helps, but the creator keeps control.
Gmail read and send permissions are powerful, so the product is intentionally approval-first. The agent can prepare a reply, but money terms, usage rights, exclusivity, legal uncertainty, and suspicious senders stay behind explicit review.
Drafts are expected to be send-ready. No bracketed contact placeholders, no fake names, no vague signoffs, and no hidden chain-of-thought in the UI. The creator sees practical summaries, tool events, outputs, and final recommendations.
Architecture
A real backend, not a hardcoded demo.
The frontend remains the creator workspace: authentication, dashboard, inbox, drafts, deals, profile, and review flows. The backend is the agent runtime. It verifies the Supabase user, reads Gmail through the user's OAuth grant, runs the CrewAI workflow, streams events, persists outputs, and writes memory back into Supabase.
The end goal
Give creators manager-level deal judgment before they hire a manager.
Real Deal is not trying to make creators less involved in their business. It is trying to make every decision clearer: which emails matter, which ones are risky, what terms are missing, how to price the reply, and what to remember for the next deal.