Deep technical explorations of software architecture, AI systems, networking protocols, and the engineering decisions that power the systems we depend on every day. Written by Barnabas Waweru.

CDN topology, Fluid compute concurrency model, request routing, build pipeline, environment isolation, function duration limits, and where cost accumulates.

Model family, generateContent vs streaming, tool use loops, File API for large media, Imagen 4 image generation, rate tier limits, and where cost accumulates.

Official xAI docs only. A Bot is one persistent named teammate. All Bots share one user-scoped cloud computer. Connectors versus computer use, approvals, skills, and routines.

AES-256-GCM, envelope keys, Argon2id, and TLS. The four encryption decisions every backend makes and how IV reuse, unauthenticated modes, and timing leaks break them.

OAuth token lifecycle, STK Push async flow, callback race conditions and idempotency, C2B vs B2C transaction types, security model without webhook signatures, and daily reconciliation.

Waiting workers, no-store version probes, Chromium-only install prompts, and Isolated Web Apps as a signed package, not a PWA flag.

24-hour session window, three-category templates, queue-first webhooks, WhatsApp Flows, and July 2025 per-message pricing.

Messages API internals, client-server tool loops, prompt caching economics, model family routing, and the Managed Agents stack.

Query engine on a copy-on-write object store: LSN branching, scale-to-zero, PgBouncer, serverless driver, read replicas.

Pricing model with Class A/B operations, presigned URL flows for private media, Workers binding pattern, CORS for browser uploads, and S3 compatibility gaps.

V8 isolate model, binding graph, Durable Objects for stateful coordination, storage tradeoffs across KV/D1/R2/DO, Smart Placement, and cost reality.

Session tokens, clerkMiddleware request interception, server-side auth helpers, Organizations multi-tenancy, and webhook sync for Next.js App Router apps.

Five caching layers, one request. HTTP Cache-Control headers, Next.js data cache, Vercel CDN, and Upstash Redis: where each layer helps and where each one bites.

Logs, metrics, and traces as one control plane. OpenTelemetry signals, W3C trace context, cardinality rules, SLI/SLO burn alerts, and redaction at emit time.

At-least-once delivery, the transactional outbox, consumer idempotency, poison messages, CloudEvents envelopes, and lag as a first-class SLO.

RFC 9110 methods, status classes, idempotency keys, problem details, versioning, cursor pagination, and the Cache-Control rules that make a public API.

Recursive resolution, DNSSEC, DoT/DoH, the TLS 1.3 handshake, certificate automation, and the failure modes that take production down while every pod stays green.

ReAct at the protocol level, client vs server tools, four memory layers, ToolLoopAgent vs HarnessAgent, loop control, guardrails, and real cost.

Event triggers, workflow orchestration, runner infrastructure, secrets management, and advanced deployment patterns for production CI/CD pipelines.

Serverless Postgres vs Cloudflare’s data platform — architecture, storage, functions, and when to pick each plane for your stack.

Service workers, caching strategies, installability, and PWA vs native tradeoffs for production web apps.

DeepSystems No. 001 — Safekeepers, Pageserver, branching, scale-to-zero, autoscaling, and connection architecture for serverless Postgres (interactive).

DeepDeepest No. 003 — foundations, taxonomy, neural nets, training lifecycle, production serving, agent tooling, and the full ML stack.

From physical signals to encrypted HTTPS packets. Firewalls, DNS resolution, TLS handshakes, IP addressing, BGP routing, network adapters, and the full packet journey traced end to end.

Inside the open-source, locally-hosted AI agent platform. Gateway control plane, ReAct agentic loop, context assembly, Skills system, memory architecture, tool execution engine, and security model.