Architectural deep dives,
postmortems & systems.
Practical lessons, architectural trade-offs, and honest reflections on designing AI systems, building scalable products, and shipping software that works in the real world.
Architecting a WhatsApp AI Receptionist: Multi-Turn Memory, 24hr Conversation Windows, and Fail-Safe Human Handoff in Production
How we bridge the official WhatsApp Cloud API to a conversational AI runtime: debounced message buffers, CRM-aware context, session memory, and a 60-second human handoff fallback.
Engineering Sub-10-Second Tactical AI on WhatsApp
How Panda FPL architectures real-time WhatsApp AI using in-memory voice note transcription, Firebase debounce locks, squad masking, and sub-10s responses.
How to Revive Dormant Clients on WhatsApp
Win back inactive clients on WhatsApp: the right timing for each service, polite one-touch messages, real message costs, and how to track the revenue returned.
Ditching Vercel for a $15 Hetzner Server with Dokploy
How we architected multi-tenant data isolation with Payload CMS and migrated from Vercel to a self-hosted Hetzner VPS and Dokploy behind Cloudflare.
How We Built an Automated WhatsApp Reputation Engine
How Patofy automates reputation on WhatsApp: queue-paced review requests, anti-fatigue cooldowns, sentiment-gated review links, dispute dampening, and Edge-rendered social proof cards.
Why Generic AI Architecture Fails in High-Context Real-Time Domains
How RAG, context management, tool architecture, and conversation memory fundamentally diverge when building agents for fast-moving, high-stakes environments.
Preventing Runaway LLM Costs with Atomic Credit Metering
How Panda FPL prevents runaway LLM costs on WhatsApp using atomic pre-flight credit deduction, automated compensation hooks, and M-Pesa STK Push.
How Consolidating 22 Tools into 5 Fixed Our Agent Hallucinations
Why exposing LLMs to 22 granular tools caused decision paralysis, and how consolidating into 5 high-order actions cut latency by half and raised tool accuracy.
Behind the Multi-Agent Architecture of Panda FPL
Why monolithic prompts fail on sports analytics, and how Panda FPL combines Swahili grit, multi-agent routing, RAG, and ClubElo data on WhatsApp.
Rebuilding Kenya's Missing Persons Network with Civil Trace
How Civil Trace re-architects civic public safety in Kenya using verified Police OB numbers, offline-first PWA resilience, and zero-cash gamification.