Triqai case study

AI-powered transaction enrichment API that dissects raw banking strings into confidence-scored merchant, location, and intermediary data. Built solo from weekend MVP to production SaaS on Cloudflare Workers.

Solo founder & engineer — Personal. Stack: Cloudflare Workers, Hono, Neon Postgres, Drizzle ORM, Next.js, React, Tailwind CSS, Stripe. Tasks: Architecture, Backend development, Frontend development, Infrastructure, DevOps. ## Origin Triqai started as a feature inside [Rediant](/works/rediant). I needed structured, categorized data extracted from raw banking transaction strings: the cryptic text your bank shows you, like `PP*DOORDASH*Convivo #1422 Santa Barbara, CA` or `AMZN Mktp US*RT4KZ0FP3`. These strings are packed with useful information but no bank cleanly parses all of it. After a weekend building an MVP with a different approach to the problem, the results surprised me. The architecture worked well enough to justify turning it into a standalone product, so I built Triqai into a full SaaS. ## What It Does Triqai is a transaction enrichment API. Send it a raw banking string, and it returns structured data: the merchant with verified business details, the physical location with coordinates, any intermediaries (payment processors, delivery platforms), and a category. Each with a confidence score indicating reliability. Unlike services that rely on static lookup databases, Triqai verifies entities against live web data. A new restaurant that opened last week has no entry in any static database, but it has a website and a Google Maps listing. Triqai finds and cross-references these sources, producing verified data for merchants that lookup-table approaches would miss. Confidence scoring is a core differentiator: rather than returning a single "best guess" without context, every entity carries a calibrated score so consumers can set their own acceptance thresholds, accept everything above 70 for a dashboard, but require 90+ before auto-categorizing a transaction in a financial product. ## Infrastructure The backend runs on **Cloudflare Workers** with no traditional servers. The architecture uses Workers for the core API and supporting services, **Durable Objects** for distributed coordination and billing atomicity, and **KV** for multi-layer edge caching. The system caches aggressively to minimize redundant expensive operations, with a distributed singleflight mechanism preventing duplicate work across concurrent requests. ## Performance All numbers measured against the live production API. The full pipeline runs genuine verification (AI inference, web searches, and cross-referencing) rather than quick database lookups. A multi-tier cache architecture turns the full pipeline into a fast cached read with a 96% latency reduction. Duplicate concurrent requests are coalesced so only one runs the full pipeline while others wait and read from cache. ## Credits & Billing Each organization gets its own billing instance acting as a single-writer credit gate with atomic checks and rate limiting. Usage counters batch in memory and flush periodically, eliminating per-request database writes on the hot path. ## Database **Neon Postgres** (serverless) with **Drizzle ORM** for type-safe schema and migrations. Multi-tenant with organization-scoped isolation. Enriched entities are shared globally (one Netflix entity benefits all users) with application-layer access control. GDPR-compliant retention with automatic data expiration and audit trails. ## Frontend & Dashboard The **Next.js** and **React** frontend includes an interactive playground to try the API before signing up, a dashboard with API key management, usage analytics, activity logs, and team management, Stripe-powered billing with plan management, and authentication with email verification, password reset, and organization invites. ## API Design The RESTful API is built with **Hono** and features a layered middleware stack handling connection management, security headers, rate limiting, API key authentication, and analytics. All running at the edge. ## Key Decisions - **Confidence scoring on every entity.** Rather than returning binary results, every entity carries a calibrated confidence score so consumers can set their own acceptance thresholds. - **Web-first verification over static databases.** Using live web data instead of a proprietary lookup database means new and small merchants are enrichable from day one. - **Distributed coordination for billing.** Using Durable Objects as single-writer credit gates eliminated race conditions in billing. ## Challenges - Designing a credit system that correctly handles deduction, overage billing, period resets, and plan changes. - Building cancellation propagation across deeply nested async operations. - Managing serverless database connection pressure at scale. - Handling the long tail of international transaction formats. - Calibrating confidence scores across different enricher types with different signal qualities.

Works

Explore selected projects with timelines, stack choices, and delivery outcomes. This view includes company and freelance examples in a horizontal timeline display.

Project List

Timeline (From 2024)

Projects

202420252026

Triqai
Today
Rediant
Through the Frames
Hytura
Knitted
Bysidde

Loading workspace