Ask Jayraj: AI Resume Assistant

OVERVIEW

A public AI assistant that answers recruiters' questions about my experience, skills and availability. Built end to end in Oracle APEX and PL/SQL on Oracle AI Database 26ai, using the Gemini API for language and embeddings. Runs at zero cost on apex.oracle.com's free tier.

THE PROBLEM

Recruiters skim a resume in seconds and often have one specific question ("Has he led a team?", "What's his notice period?"). A generic AI chatbot would invent answers; this one must only use verified facts.

HOW IT WORKS

1. AskThe visitor asks a question in any wording, with typos or in another language.
2. RewriteGemini cleans the question (fixes spelling, expands abbreviations, translates); meaningless input is detected and the visitor is asked to rephrase.
3. EmbedThe question becomes a 768-dimension vector (Gemini embedding model).
4. Vector searchOracle AI Database 26ai finds the closest profile chunks by cosine distance (top 8, with a relevance cut-off) in a native VECTOR column.
5. Grounded answerGemini writes the reply using only those chunks; if they don't contain the answer, the assistant says it doesn't know and points to my email.
6. LoggingEvery question, rewritten query, chunks used, response time and feedback is stored.

Questions about availability, notice period or salary are detected after the rewrite step and answered with a fixed redirect to email/LinkedIn, so no search or AI call is made.

KNOWLEDGE BASE · SYNCED AUTOMATICALLY FROM THIS WEBSITE

My website is the single source of truth. It exposes my CV at a private endpoint (/api/profile, protected by an API key) as one self-contained fact per line, plus a SHA-256 fingerprint header. In Oracle, a scheduled PL/SQL job runs daily:

1. Fetches the profile using an APEX Web Credential (the key is never in code)
2. Validates it: HTTP 200, plain-text content, valid fingerprint, minimum line count, and content rules (no phone numbers, no restricted topics); a bad fetch never touches the knowledge base
3. Compares the fingerprint with the last sync; if unchanged, it stops with no AI calls (about 0.2 seconds)
4. If changed, it replaces the CV document and re-embeds it (34 facts in about 25 seconds) in a single transaction, so visitors never see a half-built knowledge base and a failure rolls back automatically
5. Logs every run as UNCHANGED, UPDATED or FAILED with the reason

When I update this website, the assistant updates itself, with no manual data entry. A small manual FAQ holds only redirects (availability, notice period and salary go to email/LinkedIn).

GUARDRAILS AND ENGINEERING

• Answers only from retrieved sources; an explicit "I don't know" instead of guessing
• Prompt-injection protection in the system rules
• API key stored encrypted in an APEX Web Credential, never in code
• Retry with backoff on 429/500/503, plus an automatic fallback model
• Rate limiting in PL/SQL: per-visitor cooldown, per-session and site-wide daily caps, checked before any API call
• Output rendered as plain text (XSS-safe); JSON built with JSON_OBJECT_T
• Autonomous-transaction logging; live stats and thumbs up/down feedback on every answer
• Daily automated knowledge-base sync with change detection (SHA-256) and content validation
• Transactional rebuild with automatic rollback; every sync run logged
• Source labels taken from each fact (e.g. "Experience: Medical Guardian", "Technical Skills")
• Availability, notice-period and salary questions redirected in code, never answered by the AI
• Answers in the visitor's language

INTERESTING PROBLEMS I SOLVED

TRY IT

Not loading? Open the assistant in a new tab

TECH STACK