David Soldat #00Software engineer, Banja Luka

Green flag: open to backend and full-stack roles

Back to the pit wall
2SEN

Sentra

Codebase intelligence for GitHub repos: ask questions about any indexed repo in plain language, and have pull requests reviewed by four specialist AI agents in parallel.

View code on GitHub

Setup sheet

How the car is built, part by part
Power unit
Java and Spring Boot
Bodywork
Next.js and TypeScript with Zustand
AI package
langchain4j over Anthropic and OpenAI, four models in two bands
Fuel system
PostgreSQL with pgvector, Flyway migrations
Pit crew
Four review agents running concurrently via CompletableFuture
Radio
Hand-rolled SSE client streaming answers token by token
Paddock pass
GitHub OAuth2 sign-in, Paddle subscriptions with verified webhooks
Garage
Docker Compose

Car layout

Frontend appServiceML or AI modelData storeExternal
GitHubOAuth2, repos, PR diffsNext.js appChat, reviews, billingSpring Boot APIQuotas and model routingPaddleSubscriptionsIngestionTarball, chunk, embedRAG chatTop 20 chunks, SSEReview orchestrator4 agents in parallelMiniLM embeddingsRuns locallyPostgreSQL + pgvectorChunks and vectorsAnthropic, OpenAIEfficient and premiumsign intarballbillingretrieve
  • Next.js app to Spring Boot API
  • Next.js app to GitHub, sign in
  • GitHub to Ingestion, tarball
  • Spring Boot API to Ingestion
  • Spring Boot API to RAG chat
  • Spring Boot API to Review orchestrator
  • Spring Boot API to Paddle, billing
  • Ingestion to MiniLM embeddings
  • Ingestion to PostgreSQL + pgvector
  • RAG chat to PostgreSQL + pgvector, retrieve
  • RAG chat to Anthropic, OpenAI
  • Review orchestrator to Anthropic, OpenAI

Lap breakdown

  1. Sector 1, Data and AI

    Answers that cite file and line

    Repos are fetched as one tarball, chunked into roughly 400-token windows that keep their line numbers, embedded locally with MiniLM and stored in pgvector. Each question pulls the top 20 matches above a 0.3 similarity floor, so answers cite the exact file and lines.

  2. Sector 2, Architecture

    Four reviewers, one diff

    Security, Architecture, Performance and Docs agents share one base implementation with their own prompts and run in parallel against the same diff and codebase context. Each tracks its own status, so a failure reruns only that agent.

  3. Sector 3, Delivery and integrations

    Provider-agnostic and streamed

    Built on langchain4j's ChatModel interfaces, so nothing branches on the provider. Free users get the efficient models, Pro users can pick premium ones, and quotas use an atomic check-and-increment. Answers and reviews stream token by token over SSE.

Onboard footage

Screenshots from the app
Codebase Q&A
PR review from four agents
Usage, billing and model choice