Do things that don't scale.

Focus on manual, labor intensive tasks in the early stages, even though they wouldn't be sustainable as we grow. Prioritize building a strong foundation through direct customer interaction (TB, EXFY, DB), even if it's not scalable in the long run.

BlameGPT

BlameGPT finds the PR causing a production issue so you can go back to shipping. To prevent issues from happening in the first place, BlameGPT does a code review on your PRs.

How it started?

See this document.

How to run locally?

1. Use python version 3.13
2. Create a virtual env python3 -m venv venv
3. Activate the virtual env source venv/bin/activate
4. Install requirements pip install -r requirements.txt
5. Copy the env file cp .env.example .env, and set your personal github token, set openai api key

Warning

Do NOT set ENVIRONMENT=production in your .env file - this will create real GitHub comments on live repositories.

6. Start the server using uvicorn main:app --reload

Github token

• In your personal account, create a new token of type classic
• Scope: repo, and notification

Notebooks

To run the notebooks run jupyter lab.

Database Migrations

We use Yoyo-migrations to manage database schema changes.

• All migrations are written in pure SQL.
• The app automatically applies migrations on start. You can also manually apply them using migrate.sh
• The migrations are in libs/sqlite/[db_name]/migrations/

Naming convention

000001_init_schema.sql
000002_add_pull_requests_table.sql
000003_add_index_on_issues.sql

Flows

Code review

Automated code review triggered by @blamegpt mentions on pull requests. Reviews code changes for quality, bugs, performance, and security issues, then posts structured feedback as GitHub PR reviews with line-by-line comments.

Blame pipeline

When a new issue is created, blame gets all relevant pull requests, and performs RAG over the issue description and pull requests to rank them.

Example

Input: Issue

Users cannot login. I went to login page, submitted username and password, but I'm getting a server error.

Pipeline:

• Converts the issue to an embedding
• Gets all the recent PRs
• Converts PR title, description and code summary to embeddings, saves them to DB so new issues don't have to recompute.
• And then finds the relevant PRs using cosine similarity between the issue embedding and PR embedding.
• ~Top 20 similar PRs are sent to LLM to find the top 3 culprit PRs.
• Using blamegpt's personal github token, we add a comment to the issue.