Projects
RL environments and agent evals, backed by the distributed-systems work that makes them realistic.
DevOps Gym
RL EnvironmentsA suite of Dockerized, reproducible DevOps tasks for training and evaluating AI agents. Each task ships a realistic broken state (failing CI pipeline, broken K8s deployment, crashing service), a task spec, a hidden test-based verifier, and a difficulty tier — the building blocks of RL-environment work.
Key Features
- Dockerized, reproducible task sandboxes
- Hidden test-based verifiers per task
- Difficulty tiers and task specs
- Reward-hacking-resistant grading
K8s Debug Evals
EvalsA reproducible eval harness for measuring frontier models on Kubernetes debugging. Runs agents against broken clusters, scores pass@1 with hidden verifiers, and classifies failures into a taxonomy — built to demonstrate rigorous, non-vibe-coded eval methodology.
Key Features
- pass@1 scoring against hidden verifiers
- Failure taxonomy across models
- Reward-hacking detection examples
- Fully reproducible runs via Docker + kind
K8s Job Scheduler
Distributed SystemsA Go-based HTTP API server for prioritizing and submitting Kubernetes jobs. Features a max-heap priority queue, concurrency control, and graceful shutdown handling — the orchestration depth that makes my DevOps environments realistic.
Key Features
- Priority queue with max-heap implementation
- Kubernetes client-go integration
- Concurrency control mechanisms
- Graceful shutdown handling
HybridRAG
LLM SystemsA cost-optimized enterprise search engine with a three-layer architecture: Ingestion Pipeline for async document processing, Hybrid Retrieval combining semantic and keyword search with reranking, and a Cost Router for intelligent model selection — LLM-systems work end to end.
Key Features
- Async document ingestion pipeline
- Semantic + keyword search with reranking
- Cost-aware model routing
- Production-ready architecture
Vector Database
FundamentalsA complete vector database implementation from scratch in Python. Supports multiple distance metrics (Euclidean, Cosine, Dot Product), CRUD operations, persistence, and semantic search — fundamentals, built by hand.
Key Features
- Multiple distance metrics
- Top-k similarity search
- Metadata storage and retrieval
- Semantic search integration
Why environments, not demos?
The known failure mode of the RL-environment market is vibe-coded tasks with graders that agents can trick. Everything here is built the other way: reproducible sandboxes, hidden verifiers, explicit difficulty calibration, and documented reward-hacking resistance. Building systems from scratch — vector databases, schedulers, RAG pipelines — is how I learned what rigor looks like.