Kiragu MainaField notes
blog.alkenacode.dev / field notes Public archive

Notes from systems that have to work.

Payment rails, GPON hardware, AI infrastructure and product engineering, written from inside the build. Real constraints, real failures, real fixes.

LatestOctober 26, 2026
Field note
Read field note

Archive

Newest first
September 20, 2026
Field note

We Put Two Jev-Inspired Local AI Models in Front of Risky Software Actions

Jev introduced a new interface for AI decisions. We tested two open local interpretations, Kev and SemIf, across 202 decisions, then built a Next.js AI Bouncer around the result.

September 2, 2026
Field note

Bringing a subscriber's ONU online

From fibre in the air to a live PPPoE session with zero SmartOLT involvement: what each layer refused to do on the way.

August 29, 2026
Field note

Four webhooks, one ledger row

Three layers of idempotency across eight payment rails: a namespaced key in Redis, a unique receipt in Postgres, and an outbox that survives the crash after commit.

July 8, 2026
Dev log

Nano-banana has one identity problem

It never learns what you look like. Here is why MyPhotoAI is training per-user identity LoRAs on RunPod instead.

September 24, 2025
Benchmark

Kimi K3: Cost and Performance Edge

Kimi K3 offers strong performance at a fraction of the cost of leading American models, excelling in frontend work while earning praise from users.