03 / Lab · Side projects

What I built, what I learned.

02 case studies · AI SaaS products · Updated Apr 2026

Two live AI SaaS products, built end to end. These are not tutorials or toy apps — they are production systems that taught me how to deal with real users, LLM reliability, payments, background jobs, and the messy gap between "it works" and "it ships."

Plaudera.

plaudera.com → · SaaS · Live

AI-powered customer feedback platform for collecting product ideas, detecting duplicates, scoring demand confidence, running polls, and managing public roadmaps.

Highlights03 / many

01 Problem-first feedback system: public boards, embeddable widgets, contributor auth, voting, polls, strategic tags, and transparent roadmap decisions.
02 AI-powered prioritization: Google Gemini + embeddings detect duplicate ideas, generate similarity scores, and support confidence scoring based on votes, contributor diversity, frequency, and impact.
03 Production SaaS architecture: Next.js 16 App Router, Drizzle/PostgreSQL with pgvector, Better Auth, Polar subscriptions, Inngest background jobs, rate limits, CORS/CSRF protections, and email automation.

Stack20 items

Next.js 16React 19TypeScriptTailwind CSS 4shadcn/uiDrizzle ORMPostgreSQL +13

What I learned

Feedback quality matters more than vote count. Confidence scoring needs diversity, impact, frequency, and concentration checks to avoid building for a loud minority.
AI duplicate detection is only useful when paired with a review workflow. Embeddings find candidates, but merge/dismiss flows and audit trails make it trustworthy.
Public widgets are deceptively complex. CORS, CSRF, contributor sessions, rate limits, page rules, and bot protection become core product infrastructure very quickly.

Duebase.

duebase.com → · SaaS · Live

AI-powered financial due diligence SaaS that transforms UK Companies House filings into actionable risk assessments and financial health scores in 30 seconds.

Highlights03 / many

01 AI-powered document analysis: Google Gemini 2.5 Flash with structured JSON schema outputs processes PDF financial statements, extracting 40+ financial metrics with comprehensive risk scoring (financial health 1-5, liquidity/solvency risk levels, profitability trends)
02 Event-driven architecture with Inngest: Multi-step async background jobs handle document processing with automatic retries, real-time status updates via pub/sub channels, and granular token usage tracking for cost optimization
03 Companies House API integration: Real-time search across 15M+ UK businesses, automated data synchronization with complex mapper pattern, filing history access, and officer/PSC data enrichment with parallel processing

Stack20 items

Next.js 14 (App Router)TypeScriptMaterial-UI (MUI) v7React Query (TanStack Query v5)React Hook FormZodNext.js API Routes +13

What I learned

Gemini 2.5 Flash with structured outputs is shockingly good at financial document parsing — and an order of magnitude cheaper than GPT-4o for this workload.
Payment reconciliation is its own product. Two providers, three race conditions, one cron job. Worth getting right before you have paying users, not after.
Trial systems live or die by lifecycle logic. The 14-day reset cost me a week of bugs that only production traffic could've surfaced.