Jerry — Public AI Assistant.

A retrieval-grounded assistant that answers questions about my work in real time — built on a vector-search backend, streamed token-by-token, and tuned to never hallucinate a project that doesn't exist.

Visit live site View source

RoleSolo — design & build

TimelineMar — Jun 2025

StatusLive in production

RecognitionFeatured project · 12k+ conversations served

Addressexample.com/jerry

Built withReact · TypeScript · Supabase · pgvector · OpenAI · Edge Functions

Jerry answering a question about this very portfolio, with grounded citations.

01Overview

Most portfolio sites are static. Jerry isn't — it's a small AI assistant that lives on my site and answers visitors' questions about my projects, my stack, and how I work, in natural language.

If it can't cite it, it won't say it.

Under the hood, content is chunked, embedded, and stored in pgvector on Supabase, then streamed back token-by-token over an edge function.

02Impact

Numbers after ten weeks live

12kConversations servedsince launch

98%Answers with a citationgrounded responses

340msMedian time to first tokenp50 latency

4.8/5Helpfulness ratingn = 1,240 votes

03The interface

Fig 1Answers stream in live; each cited source links back to the original project page.

04How it came together

Ten weeks, four milestones

Week 1 · March
Indexing pipeline
Chunked every project page, embedded them, and loaded vectors into pgvector.
Week 3 · April
Grounded answering
Wired retrieval into the prompt with strict citation rules and an honest fallback.
Week 6 · May
Streaming UI
Token-by-token streaming over an edge function. TTFT under 350ms.
Week 10 · June
Public launch & tuning
Shipped to production with per-answer helpfulness voting.

05Retrieval core

Grounding every answer

retrieve.tsTypeScript

1// rank chunks by cosine similarity, then ground the prompt2export async function retrieve(query: string) {3  const embedding = await embed(query);4  const { data } = await supabase.rpc("match_chunks", {5    query_embedding: embedding,6    match_threshold: 0.78,7    match_count: 6,8  });9 10  if (!data?.length) return { grounded: false, sources: [] };11  return { grounded: true, sources: data };12}