Project · 02 / 272026

Lucky

A portfolio AI guide with a structured semantic knowledge layer, grounded answers and server-enforced security and spending controls.

View the build

Year: 2026
Role: Product direction · AI-assisted engineering · QA
Discipline: AI portfolio guide
Stack: Astro · TypeScript · Supabase · PostgreSQL · BM25 retrieval · Luna via OpenCode

01 · The project

A small companion. A carefully engineered AI system.

Lucky turns a published portfolio into a conversational guide. The work behind the mascot spans structured knowledge, fast retrieval, grounded generation, privacy boundaries, abuse controls and a disciplined release process.

TypedKnowledge preserves roles, status, dates, qualifiers and provenance

1At most one generation call per uncached typed question

EnforcedPermissions, rate limits and spending controlled by the server

Inside the system

A question. Five considered steps.

From published content to a grounded reply, with access, abuse and spending controls enforced around the conversation.

01
Supabase
Published content
Projects and approved profile facts, from one authoritative source.
02
Knowledge
Meaning and context
Typed roles, dates and qualifiers, with provenance and source relationships.
03
Retrieval
Relevant evidence
Compact coverage, BM25 retrieval and validated follow-up context.
04
Luna
A fresh answer
One bounded generation call interprets the question and writes the reply.
05
Validation
Checked references
Evidence references are checked and links use canonical portfolio destinations.

Server enforced

Access & scope: Approved public context · session checks
Abuse controls: Visitor and shared usage limits
Spending admission: Atomic quota · bounded calls

02 · Build record

What moved the work forward.

A public-facing AI product

Bill directed Lucky’s product, design and release decisions, working with Codex and Astra on implementation and testing the result in the browser. The challenge was to keep a friendly, minimal interface while handling the realities of an anonymous public website: incomplete questions, unreliable upstream answers, private information requests and limited credits.

A semantic knowledge layer

Published projects and approved profile entries become a structured knowledge snapshot. Roles, individual contribution, qualifiers, provenance and source relationships travel with the evidence. Employment and opportunity availability use typed facts with dates and freshness rules, so being employed and open to freelance work can coexist. Unspecified values remain explicit unknowns in the data model. Supabase remains authoritative and the knowledge snapshot is a validated projection of its published content.

Coverage plus fast local retrieval

Early keyword routing confused impressive-work questions with employment questions and treated some greetings as missing evidence. Lucky now gives Luna compact coverage of the published subjects, supplemented by BM25 retrieval for relevant detail. Local retrieval helps select context; the model interprets the question. Semantic structure preserves the meaning and constraints of each fact; BM25 provides fast, reproducible lexical evidence selection.

Grounded generation and a response contract

An admitted, uncached question uses one bounded Luna generation call. The answer contract carries selected evidence alongside the prose. The application checks that the references resolve to current source material and constructs links from canonical portfolio destinations. Citation handling was revised when ordinary answers failed on formatting: the source checks stayed strict while the user-facing answer became natural. Valid quotations establish traceability; they do not independently prove every generated sentence.

Follow-ups with validated context

Short follow-ups carry validated public subject context into the next question. They can stay with the previous topic or move to a newly named one, without turning arbitrary browser-supplied text into trusted knowledge. Role and contribution boundaries remain part of the answer context. When a requested detail is absent, the system can return an explicit no-support result instead of inventing it.

Guardrails outside the model

Origin and session checks, bounded requests, constrained output and server-controlled permissions protect the application around the conversation. The model receives approved public context and has no tools for reading environment files, changing policy or taking account actions. User instructions cannot grant new access. Unsafe or invalid provider output is withheld by application checks.

Cost and abuse controls

Spending is admitted atomically in the database before a paid attempt. Per-visitor limits, shared allowances and concurrency controls bound usage across requests. Generation size is capped and an upstream failure does not trigger an automatic paid retry. Complete greetings and authored shortcuts avoid generation. These choices keep cost control independent of a model’s willingness to refuse a request; limits bound consumption without promising uninterrupted availability under abuse.

Privacy and content boundaries

Generated-answer caching is encrypted and isolated by session, with cached answers checked again against their evidence. This avoids reusing one visitor’s generated wording in another visitor’s session. Published knowledge stays separate from private operational information and from the authored Info page. Splitting those boundaries also stopped raw knowledge notes from becoming public editorial copy.

Reliability through failure diagnosis

Timeouts, evidence rejection and usage denials once produced the same generic fallback. Distinguishing those stages made debugging useful, while stricter evidence-contract handling removed citation IDs and avoidable failures from ordinary answers. Provider calls retain a bounded deadline and failure path. Lucky can explain that an answer is unavailable; it does not fill the gap with a guessed fact or quietly spend on retries.

Tested behaviour and deliberate delivery

The protected beta was refined through Bill’s testing and feedback from friends. Recorded examples cover ordinary answers and follow-ups; fresh local Luna tests cover private-information refusal, an unknown detail and a false premise. The reviewed beta release passed 297 tests across the portfolio repository. Local review, protected beta testing and main-site publication remain separate release decisions, with a specific build reviewed before promotion.

In conversation

Real answers. Real boundaries.

Everyday questions, follow-ups and boundaries, shown through six observed replies. Replays are labelled; the hobbies image is an original beta screenshot.

01 / Guardrail · Private informationResponse replay