Llm Architect
LLM system architect — model selection with empirical benchmarking, fine-tuning strategy, inference optimization, evaluation framework design, and production system architecture for LLM-powered applications.
Nlp Engineer
NLP engineer for text processing, classification, NER, embeddings, and information extraction — using the right tool for each task from regex to fine-tuned transformers.
Recommendation Engine
Recommendation systems engineer — collaborative filtering, content-based methods, hybrid architectures, two-stage retrieval/ranking, cold-start handling, and A/B testing for personalization systems.
Devops Infra
Infrastructure engineer for containerization, Kubernetes deployment, and CI/CD pipelines — Docker, docker-compose, K8s manifests, Helm charts, GitHub Actions, and GitLab CI.
Computer Vision Engineer
Computer vision engineer for image classification, object detection, segmentation, and video analysis — from dataset curation through optimized production inference.
Data Scientist
Statistical analysis, EDA, hypothesis testing, and reproducible insight generation from datasets. Prioritizes scientific rigor — effect sizes, confidence intervals, and causal reasoning over surface-level pattern matching.
Task Planner
Breaks an approved spec and architecture into a dependency-ordered, bite-sized task list with acceptance criteria, verification steps, and explicit checkpoints. Human reviews before any implementation starts.
Code Reviewer
Reviews diffs and pull requests against the spec and task acceptance criteria. Produces structured feedback — must-fix, should-fix, nit — with specific line references. Does not rewrite code.
Developer
General implementation agent that executes tasks from a task plan incrementally — one slice at a time, building and testing between each, committing after each verified slice. Language-agnostic; defers to language-expert for language-specific idioms.
Codeassist Guardrails
Behavioral guardrails for LLM coding derived from Andrej Karpathy's observations on AI coding pitfalls. Apply whenever writing, editing, refactoring, debugging, or reviewing code — any language, any stack. These principles should be active any time Claude touches code.
Data Engineer
Data pipeline engineer for batch and streaming workflows — ETL/ELT, Spark, data warehousing, orchestration, and data quality. Designs for idempotency, schema evolution, and observability from the start.
Diagram Definition
Generates diagram-as-code definitions in Mermaid and C4/PlantUML from an architecture document. Output is markdown-embeddable code blocks — no rendering, no external tools required.
Architect
Translates an approved spec into a technical architecture — component breakdown, data flow, API surface, technology decisions with tradeoff analysis, and identified risks. Outputs a reviewable architecture document before any implementation starts.
Ml Engineer
End-to-end ML pipeline engineer — data ingestion through model serving. Covers feature engineering, training, evaluation, and deployment with a hard focus on reproducibility and train-serving consistency.
Prompt Optimization
Prompt engineer that takes a rough idea or draft prompt and produces an optimized version by systematically applying prompt engineering techniques — chain-of-thought, few-shot, role framing, constraint injection, output structuring, and more.
Vector Database Engineer
Vector search and semantic retrieval engineer — embedding pipelines, chunking strategy, vector store selection, index configuration, hybrid search, and retrieval evaluation.
Mlops Engineer
ML infrastructure engineer for model lifecycle management — model registry, serving infrastructure, CI/CD for models, A/B testing, and production monitoring. Bridges experimentation and reliable production systems.
Spec Writer
Turns a feature one-pager or raw requirements into a structured, technology-agnostic specification — the shared contract between product and engineering before any code is written.
Skill Observer
Review pending skill suggestions collected by the background observer. Use when the user wants to see, approve, reject, or apply pattern suggestions that the observer flagged during past sessions.
Ai Scientist
Adaptive scientific persona agent — classifies the domain of the user's AI/ML request, assumes the relevant expert identity, then conducts end-to-end scientific work including hypothesis formation, experiment design, implementation, and evaluation.
Feature Brainstorm
Interactive ideation agent that sharpens a raw feature idea into a concrete one-pager — problem statement, recommended direction, MVP scope, and explicit tradeoffs — through structured divergent and convergent thinking.
Deepresearch
Web-aware research agent that classifies the domain, then executes targeted multi-source searches across arXiv, Papers With Code, GitHub, and domain-specific blogs to produce a structured synthesis with citations.
Autoresearch
ML experiment optimization agent that uses tree search to explore solution approaches — implements changes, measures against a fixed metric, keeps improvements, reverts failures.
Content Curator Research
Autonomous 3x/day AI intel sweep — fetches HackerNews, GitHub Trending, Google AI Blog, ArXiv, X/Twitter, and star-history.com, filters for AI/LLM/agent topics, appends to a daily research file in a GitHub repo with a "What's Hot Right Now" table of contents followed by deep-dive sections.
Data Visualization
Data visualization engineer — transforms datasets into accurate, accessible, and interactive charts and dashboards. Prioritizes perceptual accuracy and accessibility over visual complexity.
Mcp Developer
Builds production-quality MCP (Model Context Protocol) servers and tools from scratch using the TypeScript and Python SDKs — tool schemas, resource definitions, prompt templates, and transport configuration.
Language Expert
Adaptive language expert that classifies the language and framework from the task, assumes that expert persona, and implements using language-specific idioms, tooling, testing conventions, and best practices. Covers Python, Rails, Java, Kotlin, React, and TypeScript.
Content Curator
Draft a LinkedIn post or newsletter section grounded in real data from the latest AI intel report. Invoke when the user says "write a LinkedIn post", "draft a post about [topic]", "write a newsletter section", "curate content for me", "turn today's intel into a post", or "I want to post about [X]". Always reads the latest research-MMDDYY.md from <YOUR_LOCAL_REPO_DIR>/content-curator/ first and ci…