Azeez Roheem

I spent two weeks applying prompt engineering to a real SaaS product. Moved rewrite quality from 7.37 → 8.37/10 across 38 bullets. #buildinpublic #promptengineering #AI #SaaS

Result: PDF → tailored resume in 30 seconds. Upload CV → paste job description → get tailored bullets that match the role. Free to start → resumetailor.cv Next.js · OpenAI · pgvector · Neon · Clerk · Stripe · Upstash · Sentry

What I'd do differently: → usage_events on day 1, not month 4 Can't improve what you can't measure → Sentry before first user, not first bug → Drizzle from the start — raw SQL = tech debt → Zod on every AI response, immediately

None of this is novel. All of it is necessary. If you're shipping an AI product and finding out about bugs from users — this is the week to fix that. Full post with code snippets: azeezroheem.hashnode.dev/i-added-produc…

Day 9: /api/health + status indicator SELECT 1 against Neon + openai.models.list() in parallel → "● All systems operational" on the homepage → polls every 60s, zero tokens consumed

I had a working AI SaaS. Then a user emailed me about a bug I didn't know existed. I had no alert. No structured errors. No usage tracking. I spent 10 days fixing that. Here's what I built 🧵

@EOEboh I want to know too. @Dominus_Kelvin over to you.

As a Claude user, give me reasons I should switch to Codex

Full post — embeddings, pgvector setup, chunking strategy, before/after data, real code: [azeezroheem.hashnode.dev/how-i-added-se…] Live product: resumetailor.cv #buildinpublic #RAG #AI #SaaS

Cost per resume analysis: $0.000021 At 10,000 sessions/month: $0.21 total embedding cost. Embedding is free. Latency is the problem. Warm requests: 8–10s (all GPT, not vectors)

Your resume says "led team of 5" The JD says "managed engineers" Keyword matching: 0 overlap. Filtered out. Semantic matching: 0.48. Strong match. Same skills. Different words. I spent 2 weeks fixing this in Resume AI Tailor 🧵