CS-15 — Automation & AI · Backend & integrations · Architecture & performance

This engineering portfolio with an AI assistant

A multilingual portfolio with an experience explorer, case studies and an AI assistant that answers only from an approved knowledge base, with cost controls.

My role: Design and implementation, with AI-assisted development.
Period: 2026
Disclosure: Public project
Technologies: TypeScriptNext.jsPostgreSQLDrizzle ORMAnthropic APIVercel
Last reviewed: 13 November 2026

Context

A static CV does not show how someone reasons about technical problems. The goal was an explorable portfolio with an assistant that is useful, accurate and very cheap to run.

Challenges

01Preventing the assistant from inventing experience or metrics.
02Keeping the cost per conversation in cents and protecting it against abuse.
03Publishing in three languages without silent or unreliable translations.

Solution & architecture

Version-controlled Markdown compiled at build time (drafts never reach the bundle), deterministic context selection without a vector database, explicit per-mode routing, and atomic usage and budget limits in PostgreSQL.

This portfolio

sql

SELECT ai_reserve(
  ARRAY['visitor','global'], ARRAY[$1,'all'], ARRAY[$2,$3],
  ARRAY[15,500], ARRAY[NULL,1.00]::numeric[], ARRAY[1,NULL],
  $4::numeric
);

Atomic reservation: if any limit fails, no counter is modified.

Key decisions

D-01

No vector RAG in the first version

The knowledge base is small; sending the full approved context or a tag-based selection is simpler, deterministic and cheaper.

D-02

Budget counters in PostgreSQL

An atomic SQL function reserves budget before each call and reconciles it with actual usage, without adding another paid service.

Trade-offs

The CSP allows inline scripts to keep pages static, at the cost of not using per-request nonces.

Results

Results will be published once verified

Lessons learned

Not documented yet

Ask about this case study

Ask about this case studyAI · answers from an approved knowledge base · not Ander in real time

Questions here are answered with the case study “This engineering portfolio with an AI assistant” as context.