Case study

StellaAI Assistant for Stellar Developers

The Stellar Development Foundation's AI assistant, answering Stellar and Soroban questions on Discord, Telegram and developers.stellar.org from a knowledge base of Stellar's own docs, code and community.

View gallery15

Role: Lead Developer
Timeline: May 2025 - Mar 2026
Type: Project I worked on
Stack: TypeScript, Bun +12

Overview

Stella is the AI assistant the Stellar Development Foundation (SDF) offered developers building on Stellar and Soroban, its smart contract platform. Developers asked it questions in the #stella-help channel of the Stellar Developers Discord by starting a message with "Stella, …", through a yellow chat icon on developers.stellar.org and the Stellar Lab, and, for the Stellar Community Fund (SCF), through a Telegram bot. Its answers came from a knowledge base of Stellar's own content: the docs and SDK repositories and their GitHub issues, Stellar Stack Exchange, SDF's YouTube talks, Discord threads, the stellar.org website and the SCF handbook, and each answer cited its sources and suggested follow-up questions. SDF launched the first Stella in 2024, and Jutsu rebuilt it in 2025; I led that rebuild. I wrote most of the new backend, a Bun and Hono monorepo that ingests every source on a schedule into Qdrant, answers with Claude through the Vercel AI SDK, runs the Discord and Telegram bots and serves the website widget. I also wrote most of Stella Studio, the React app where the team manages sources, tests retrieval and prompts, reviews every conversation and rating, and measures answer quality with evaluation runs, and added YouTube embeds, follow-up questions and file uploads to the chat widget, which Hypersolid built. The studio screenshots show it running locally with made-up data, and the widget shots show the production widget on a developers.stellar.org page, answering from a local mock.

Highlights

01Stella in the Stellar Developers Discord: start a message with "Stella, " and the answer streams into a new thread, with source links, follow-up questions and Like, Dislike and Wrong buttons
02A consent step before a user's first question, a server and channel allowlist managed from the studio, and per-channel rate limits
03A chat widget on developers.stellar.org and the Stellar Lab, with threads, file attachments, a short or long answer setting and feedback on every answer
04A Telegram bot for the Stellar Community Fund that can also look up SCF projects, awards and submissions in Airtable
05Ingestion from GitHub docs and issues, Stack Exchange, YouTube transcripts, Discord channels, website sitemaps, GitBook, Notion and uploaded PDF, Word and CSV files, each on a manual, cron or interval schedule
06Retrieval over Qdrant with OpenAI embeddings, HyDE query expansion and a Claude reranker, and answers from Claude Sonnet 4.5 with inline citations
07Two projects, Stella and SCF, each with its own sources, bots and analytics
08A query playground that shows the retrieved chunks and their scores, a chat playground and the widget's embed code
09Analytics of every question, thread and rating, with topic clustering and trending topics
10Evaluation runs over question datasets, scored by token F1, semantic similarity, recall@K, LLM-judged correctness and groundedness
11Editable system prompts, API keys for the RAG service, Google sign-in limited to approved domains, and Sentry monitoring
12Deployment to AWS Fargate with SST, in separate staging and production stages

Gallery

Inside the product

15 screens · click any to enlarge

01The yellow Stella launcher on a developers.stellar.org docs page.

02The Stella widget answering a CLI question with a code block, cited docs, follow-up questions and feedback buttons.

03The studio's chat playground: deploying a Soroban contract to testnet with the stellar CLI.

04The query playground, with HyDE and reranking switched on, per-collection limits and the retrieved chunks with their scores.

05Knowledge sources: the Stellar docs and SDK repositories, GitHub issues, Stack Exchange, SDF's YouTube channel, Discord, the developers.stellar.org sitemap and the SCF Handbook.

06Ingestion jobs by source, with progress, vectors and tokens.

07Scheduled re-ingestion: each source's cron schedule, with its next and last run.

08The Discord integration: the bot's connection status and what it does in the server.

09Choosing the Discord channels Stella may answer in.

10Thread analytics: every conversation, with its question, user and time.

11Topic analysis: a word cloud of what developers ask about.

12Feedback analytics: Like, Dislike and Wrong ratings with comments, by platform.

13Evaluation runs comparing a baseline with HyDE, reranking, chunking and prompt changes.

14One evaluation run, scored by token F1, semantic similarity, LLM-judged correctness and groundedness, by question category.

15Editing Stella's default system prompt.

Next project

AgentRel | AI Assistants Grounded in Your Docs

A multi-tenant platform that turns a product's docs, GitHub repos, websites and videos into an AI assistant that answers with cited sources, in a dashboard and an embeddable chat widget.

View case study