Building an Autonomous Zero-Touch Client Onboarding Agent with Gemini 3.5 Flash & Google Cloud

The Agency Onboarding Bottleneck

At digital and web development agencies, converting an interested inbound lead into an active project usually takes days of friction-heavy back-and-forth:

Manual Qualification Calls: Spending 30–60 minutes asking foundational project scoping questions.
Custom Proposal Drafting: Writing formal Business Requirements Documents (BRDs) and scope breakdowns manually.
Pricing & Tier Negotiations: Calculating rates, factoring in custom feature add-ons, and avoiding mispriced deliverables.
Invoicing & Deposit Chasing: Generating payment invoices, sharing bank links, and constantly following up on unpaid advance deposits.

For a fast-moving agency like TaskDrift, this manual pipeline creates an expensive bottleneck.

We set out to eliminate this entire pipeline with an autonomous, zero-touch multi-agent system powered by Google Cloud and Gemini 3.5 Flash - compressing the entire intake-to-deposit cycle down to under 60 seconds.

We built this project and live demo specifically for the #AllThingsAgentic Hackathon.

The 5-Agent Autonomous Architecture

Rather than relying on a single monolithic prompt, we separated responsibilities into 5 specialized autonomous agents with strict guardrails and state machine coordination:

[ INBOUND LEAD ]
       │
       ▼
1️⃣ [ INTAKE AGENT ] ──(17 Regex Guardrails + Gemini 3.5 Flash)
       │  ↳ Sanitizes inputs, asks 1–3 contextual clarifying questions if scope is vague
       ▼
2️⃣ [ DRAFTING AGENT ] ──(Grounded BRD Generation)
       │  ↳ Compiles formal proposal locked to fixed pricing catalog invariants
       ▼
3️⃣ [ FINANCE AGENT ] ──(Scope-Creep Detection + Razorpay Integration)
       │  ↳ Detects tier mismatches, computes totals, generates 50% advance payment link
       ▼
4️⃣ [ FOLLOW-UP AGENT ] ──(Omnichannel Dispatch)
       │  ↳ Dispatches via Meta WhatsApp Cloud API with instant Resend Email fallback
       ▼
5️⃣ [ AUTONOMOUS RECONCILIATION ] ──(Cloud Scheduler + Firestore)
          ↳ Hourly worker verifies deposit, creates CRM Project record & alerts leadership

1. Intake Agent (Gemini 3.5 Flash + 17 Regex Guardrails)

The intake agent acts as the frontline gatekeeper. Incoming customer submissions are first passed through 17 regex-based security guardrails designed to detect and sanitize prompt injections, instruction overrides, and malicious payloads before reaching the model.

Once sanitized, Gemini 3.5 Flash evaluates the project requirements. If the user input is ambiguous or underspecified, the agent dynamically generates 1–3 sharp, highly contextual clarifying questions to establish explicit project boundaries.

2. Drafting Agent (Grounded BRD & Scope Definition)

Once scope details are locked, the Drafting Agent compiles a comprehensive, professional Business Requirements Document (BRD).

To eliminate AI hallucination on services and pricing:

The agent is constrained by strict pricing catalog invariants.
Every deliverable, milestone, and timeline is grounded directly in TaskDrift's predefined service schemas.
Zero invented deliverables or unauthorized discount rates.

3. Finance Agent (Scope-Creep Detection & Payment Link Generation)

Scope creep is one of the most persistent risks in software client onboarding. The Finance Agent inspects the generated BRD for feature-to-tier mismatches:

If a client selects a Starter Tier but requests e-commerce features (e.g. shopping carts, payment gateways, product inventory), the agent automatically upgrades the quote to the E-commerce Tier with clear itemized line items.
It calculates the statutory taxes and automatically generates a 50% advance Razorpay payment link configured with exponential backoff retries.

4. Follow-up Agent (Resilient Omnichannel Dispatch)

Once the proposal and payment link are generated, the Follow-up Agent handles instant multi-channel dispatch:

Primary delivery via Meta WhatsApp Cloud API directly to the client's phone for near-instant open rates.
Automatic, zero-delay fallback to Resend Email API with full HTML formatting and PDF proposal attachments if WhatsApp delivery encounters a network or webhook failure.

5. Autonomous Reconciliation (Google Cloud Scheduler & Firestore)

Closing the loop requires zero human intervention:

Google Cloud Scheduler triggers an hourly reconciliation cron job that polls payment status invariants directly from Razorpay.
The moment an advance deposit is captured and confirmed, the system autonomously creates an active CRM Project Record in Google Cloud Firestore (with local SQLite WAL fallback for extreme fault tolerance).
Immediate real-time notifications are dispatched to agency leadership via webhooks, moving the project straight to the development queue.

The Cloud & Tech Stack

We engineered the entire system for speed, isolation, and high availability:

Live Demo & Open Source

The entire platform is live on Google Cloud Run and fully open-sourced:

🌐 Live Client Portal & Dashboard: TaskDrift Onboarding Demo
💻 Open-Source GitHub Repository: github.com/mohanprasath-dev/taskdrift-onboarding-agent

Hackathon Submission

This autonomous agent system was designed and implemented for the #AllThingsAgentic Hackathon, demonstrating how multi-agent architectures on Google Cloud can replace hours of manual agency operations with an instantaneous, deterministic, and frictionless customer onboarding experience.