AI Engineer (Founding Engineer – AI)

Role Overview You will be the intelligence architect of Vibecoderz. As the AI Engineer, you’ll design, implement, and optimize the multi-agent orchestration system that powers the TutorAgent, PlannerAgent, ScreenPerceptionAgent, and CodeAgent. This is not a research-only role. It’s applied AI at scale: integrating Gemini models, Pub/Sub protocols, LangSmith evals, and Neo4j graphs into a production-grade learning platform. You’ll work closely with Backend, Prompt, and Frontend engineers to ensure AI capabilities feel seamless, fast, and trustworthy for 150M+ developers worldwide. You’ll use Linear for task management, Notion for specs and experiments, and GitHub for code reviews, making the AI system’s evolution transparent and collaborative.

Key Responsibilities

• Multi-Agent Orchestration - Implement TutorAgent and sub-agents using the Google Agent Development Kit (ADK), ensuring modularity and reliability.
• Model Integration - Integrate Gemini Pro, Gemini Vision, and Gemini Live API for multimodal reasoning (text, vision, voice).
• Communication Protocols - Design and maintain Agent-to-Agent (A2A) communication via Google Pub/Sub. Guarantee low-latency, async messaging.
• Context & Memory Systems - Build working memory in Redis, long-term user data in Firestore, and Developer Graph in Neo4j to support adaptive tutoring.
• Artifact Generation - Collaborate with Prompt Engineers to refine artifact workflows: slides, quizzes, code snippets, and runnable mini-apps.
• Evaluation & Guardrails - Build evaluation pipelines in LangSmith/Langfuse to test prompt stability, reduce hallucinations, and monitor drift.
• RAG & Vibe Browser - Integrate retrieval-augmented generation (RAG) pipelines with GitHub docs, StackOverflow, and MDN for contextual support.
• Adaptive Learning Loops - Implement performance-tracking systems that adapt course flow based on learner progress and quiz outcomes.
• Scaling & Optimization - Benchmark and optimize model selection via hybrid routers (Gemini Flash for speed, Pro/Claude for depth).
• Security & Safety - Implement guardrails against unsafe generations, prompt injection, and biased outputs.
• Cross-Team Collaboration - Translate PM requirements into AI system designs, coordinate with Backend for APIs, and FE for real-time outputs.

Success Metrics

90 Days (Probation):

• TutorAgent and PlannerAgent integrated with Gemini Pro + Pub/Sub messaging.
• Redis working memory and Firestore user schema live.
• LangSmith evaluation pipeline created with baseline metrics (<15% hallucination rate).

12 Months:

• Orchestrate 10+ specialized agents in production.
• Reduce TutorAgent response latency <3s end-to-end.
• Adaptive learning system live with >40% improvement in learner retention.
• AI drift monitoring and guardrails automated in CI/CD.

Must-Haves

• 10+ years in applied AI/ML engineering.
• Expertise in LLM orchestration frameworks (LangChain, CrewAI, ADK).
• Deep experience with multimodal model integration (text, voice, vision).
• Strong knowledge of messaging systems (Pub/Sub, Kafka, or similar).
• Proven delivery of AI-first products in production environments.

Nice-to-Haves

• Research background in NLP, RLHF, or agentic AI systems.
• Contributions to open-source AI frameworks.
• Prior work on developer-focused AI products.
• Startup/founding engineer experience.

Tech Stack Visibility

• Core AI: Google ADK, Gemini Pro, Gemini Vision, Gemini Live API
• Communication: Google Cloud Pub/Sub
• Memory: Redis (working), Firestore (archival), Neo4j Aura (Developer Graph)
• Eval & Prompting: LangSmith, Langfuse, hybrid model router
• Infra: Cloud Run, API Gateway, GitHub Actions
• Tools: Linear (execution), Notion (PRDs/experiments), GitHub (repos)

Assessment Objective: Validate ability to design and implement a production-ready multi-agent system.

Challenge (Candidate PoC):

• Build a TutorAgent that:
• Takes input: “Teach me React Hooks.”
• Delegates to sub-agents:
• CurriculumAgent → outline
• ContentAgent → lessons
• CodeAgent → runnable code snippet
• QuizAgent → quiz JSON
• Stores all results in Firestore + Developer Graph (Neo4j).
• Communicates via Pub/Sub.
• Add a Learning Adaptation Loop:
• If quiz score <60%, regenerate lesson with simplified examples.
• Evaluation & Guardrails:
• Set up LangSmith eval pipeline with at least 20 golden prompts.
• Implement guardrail filter to block unsafe outputs.
• Performance Targets:
• TutorAgent orchestration end-to-end <3s.
• Sub-agent response <1.5s each.

Deliverables:

• Multi-agent orchestration codebase.
• Firestore + Neo4j schema examples.
• Evaluation report (accuracy, latency, hallucination rate).
• GitHub repo with CI integration.
• 5-min Loom demo walkthrough.

Evaluation Criteria:

• Architecture & Orchestration Design (30%)
• Model Integration & Multimodal Handling (20%)
• Evaluation & Guardrails Implementation (20%)
• Performance & Scalability (15%)
• Documentation & Testing (15%)