Field notes from the bench.
Slow blogging. We post when we have something honest to say about AI engineering, agency economics, or what's actually working in the Pakistan-Gulf software lane. No SEO bait, no listicles, no AI-written posts about AI.
Why your RAG pipeline keeps hallucinating and a 6-step fix that actually held up in prod.
Most RAG systems we audit fail in the same three places: chunking strategy, retrieval evaluation, and the silent assumption that more context is always better. Here's the playbook we now use on every legal-tech and healthcare client built around evals, not vibes.
Read the full pieceMore to read.
Sorted by recency. Pieces tagged Engineering are deep dives;Operations is about running an agency; Markets is industry analysis.
The 5 KPIs we track on every AI voice deployment (and the 3 vendors lie about).
Concurrency, real handle time, deflection, escalation accuracy, cost-per-resolved-call. Industry "demo" metrics rarely survive contact with a live patient queue at 9am.
Pakistan IT exports hit $3.8B. What that actually means if you're starting an agency.
A look at the headline number, the segment that's actually growing, and where the price floor sits in 2026.
Cutting our OpenAI bill by 64% without losing quality.
Routing, cache layers, prompt compression and the smaller-model fallback that actually pulled its weight.
How we write SOWs that don't quietly stretch.
A template, plus three clauses we always include, and the one we removed after a hard lesson.
WhatsApp Business API: 9 gotchas we learned the hard way.
Template approval timelines, the 24-hour session rule, BSP pricing arbitrage, and why your webhook keeps timing out.
Hiring our first engineer: 60 days, 200 applications, 1 offer.
The take-home, the interview loop, and the candidate signal that turned out to matter most.
The "voice-AI agent" market is too big, and the wedge that still works.
Six-billion-dollar TAM slides are everywhere. The real question is which vertical, which channel, and which integration moat actually survives the next 18 months of platform shifts.
Multi-tenant Postgres without losing your weekends.
Schema-per-tenant vs. row-level security: the migration we ran twice, the second one done right.