shoaib /decisionsRSS JSONv2026.07.5

What I chose,and the bill it ran up.

A public log of engineering decisions made under real constraints. Each one names the call, the alternatives, the trade I made, and the trigger that would force me to revisit. Nothing here is a pattern. All of it is contingent.

JULY 20261 entry

ADR-005
11
JUL
ACCEPTEDImpact · SAI
Feed the engineering-decision log into the AI twin's system prompt
The strongest senior signal on the site — this decision log — was invisible to the AI twin. A visitor asking the chatbot about a hard trade-off got nothing. The chat now pulls the structured fields of the most recent published decisions into its system prompt and points deeper questions at /decisions.
#ai#groq#rag#portfolio

MAY 20261 entry

ADR-004
18
MAY
ACCEPTEDMOBILE
Clamp every console.log argument before it crosses the Capacitor JS↔native bridge
1,900 OOM crashes/week on low-RAM Android, all from the same Crashlytics signature, were caused by Capacitor's Android bridge running String.format on every console.log payload before any log-level guard. Fixed with a runtime guard that clamps any log argument to 8 KB before crossing the bridge, layered with build-time stripping and an agent safety rule.
#mobile#android#capacitor#performance#production-incident

APRIL 20261 entry

ADR-003
09
APR
ACCEPTEDImpact · SAI
Gate the chat with a client-side keyword filter before spending a token
The chat endpoint costs tokens and is rate-limited, and people will try to use it as a free general-purpose LLM. A cheap client-side keyword check blocks obviously off-topic questions instantly, before any network call; the server rate limit is the real backstop.
#ai#cost#rate-limiting#frontend

MARCH 20261 entry

ADR-002
22
MAR
ACCEPTEDImpact · SAI
Run the AI twin on llama-3.1-8b-instant, not 70b
The 'chat with me' twin answers short, factual questions grounded in my CMS data. An 8B model, streamed, with a rich grounded system prompt feels instant and costs almost nothing — the job is retrieval and phrasing, not open-ended reasoning.
#ai#groq#llm#performance

FEBRUARY 20261 entry

ADR-001
14
FEB
ACCEPTEDImpact · MAUTH
Disable Clerk auth entirely inside embedded browsers
Portfolio links get opened inside in-app WebViews (LinkedIn, Instagram, X) where OAuth popup/redirect flows silently break. Rather than fight the WebView, I detect it and skip Clerk entirely, serving the full public portfolio with auth-gated surfaces hidden.
#auth#clerk#webview#nextjs

Chat with Shoaib

Chat with Shoaib

What I chose,and the bill it ran up.

Feed the engineering-decision log into the AI twin's system prompt

Clamp every console.log argument before it crosses the Capacitor JS↔native bridge

Gate the chat with a client-side keyword filter before spending a token

Run the AI twin on llama-3.1-8b-instant, not 70b

Disable Clerk auth entirely inside embedded browsers