Skip to main content
All work

Consumer SaaS

Email Intelligence SaaS Platform

The hard part: Five AI providers behind one registry, structured outputs schema-checked before anything downstream trusts them.

LLM pipeline that classifies Gmail into structured life events with Zod-validated outputs. Five AI providers behind one gateway, AES-256-GCM token storage, three-tier monetisation.

  • 5 AI providers unified
  • SaaS startup
  • Australia
  • Shipped 2026
  • Still in production

5

AI providers unified

3

Scan cadences: 5-min, 2-hr, daily

AES-256

GCM server-side encryption

Email Intelligence SaaS Platform — Consumer SaaS
Photo: Faraaz Zuberi / Unsplash

How it actually runs

  • Model
  • Guardrail
  • Deterministic
  1. Inbox scan

    5-min, 2-hour and daily cadences

  2. AI Gateway routing

    5 providers, one model registry

  3. Classification

    Structured output

  4. Zod validation

    Schema-checked before it is trusted

  5. AES-256-GCM

    Server-side, no browser key exposure

  6. Supabase RLS

    Row-level isolation per tenant

The problem

A consumer SaaS needed to turn a Gmail inbox into structured, actionable life events at scale, across multiple AI providers, without leaking a single token.

The approach

We built the pipeline on Next.js App Router with Supabase RLS: LLM classification with Zod-validated structured outputs, multi-tier scanning with concurrency controls, and Vercel AI Gateway as unified routing across five providers. Auth moved from browser-passphrase E2EE to server-side AES-256-GCM.

The outcome

Emails become structured events on 5-minute, 2-hour and daily cadences. A class of client-side key exposure is gone. Free, Plus and Pro tiers earn from day one.

A consumer SaaS needed to turn a Gmail inbox into structured, actionable life events at scale, across multiple AI providers, without leaking a single token.

Pipeline

  • LLM classification with Zod-validated structured outputs
  • Multi-tier scanning with concurrency controls across 5-minute, 2-hour and daily cadences
  • Vercel AI Gateway as unified routing across five providers with a central model registry
  • Next.js App Router with React Server Components and Supabase RLS
  • PKCE-based Google OAuth

Security

We replaced a browser-passphrase E2EE model with server-side AES-256-GCM token encryption, eliminating a class of client-side key exposure vulnerabilities without changing the user experience.

5

AI providers unified

3

Scan cadences

3

SaaS tiers

Monetisation

Free, Plus and Pro tiers with scan quotas, AdSense and rewarded-ad credit top-ups. The product earns from day one instead of waiting for a pricing page.

More like this

Conversational AI Agent for Live ERP Data — Enterprise operationsEnterprise operations

NL→SQL

Plain English to live data

100%

Prompt-injection hardened

Conversational AI Agent for Live ERP Data

Every business question needed a developer, and answers arrived days late.

The hard part: Letting non-technical staff query live ERP data without ever giving the model database access.

A natural-language interface over live ERP data. Non-technical staff ask plain-English questions and get ORM-safe, role-gated answers in seconds.

  • Claude MCP over live ERP
  • Enterprise ERP client
  • Shipped 2025

Read case study

Healthcare Insurance Claims Automation — HealthcareHealthcare

100%

Claims processed automatically

80%+

Manual data entry eliminated

Healthcare Insurance Claims Automation

A Canadian health platform was keying every OHIP claim, eligibility check and reconciliation by hand.

The hard part: Three government interfaces — OHIP, MCEDT and EDI — each with its own failure modes, running unattended overnight.

OHIP real-time eligibility checks, MCEDT payment reconciliation and automated XML/EDI batch claims. What took a team days now runs overnight without oversight.

  • OHIP + MCEDT
  • Runs overnight, unattended
  • Healthcare platform

Read case study

Forecasting & Traceability Dashboards — Distribution & inventoryDistribution & inventory

+45%

Inventory planning accuracy

−60%

Stock discrepancies

Forecasting & Traceability Dashboards

Multi-warehouse inventory planned on last month’s numbers, with discrepancies found only at count time.

The hard part: Forecasting that stays honest when the underlying stock data has been wrong for years.

Time-series sales forecasting and end-to-end stock traceability dashboards. Inventory planning accuracy up 45%, stock discrepancies down 60%.

  • Multi-warehouse
  • Multi-warehouse distributor
  • India

Read case study

Next step

Discuss a similar build.

Tell us what your version of this problem looks like. The first 45 minutes are free and specific.

Straight to the engineer. No form, no gatekeeper.

Straight to the engineerReply in 4 business hoursNDA on request

Accepting new engagements · Q4 2026