// stack / tools / rankings
Tech stack & AI tools
Disclaimer. Rankings here are my personal opinions — not universal truth or a live snapshot of the current state of the field. I still review anything that touches production; models speed exploration but do not replace ownership.
My State of AI
Models
Composer 2
Composer 2 is my workhorse in Cursor for UI and day-to-day implementation. For the price, speed, and capability, it is fantastic for layout, component iteration, visual cleanup, and keeping momentum in the repo.
+ Pros
- Outstanding for UI work and design-in-the-editor iteration
- Excellent value-to-speed ratio for day-to-day product work
− Cons
- Tied to Cursor subscriptions and routing
- Less suited than Opus for ambiguous deep research and hardest reasoning
Claude Opus 4.6
Opus 4.6 is my heavy-focus model — where I go for deep research, long-form technical discussions, and anything that demands sustained reasoning. It is the model I trust most when the task is messy, the architecture is unclear, or I want a serious thinking partner on difficult engineering tradeoffs. If I need to go deep, I open Claude.
+ Pros
- Best-in-class for deep research and long-form technical discussion
- Excellent for architecture reasoning, critique, and problem framing
- The model I trust most for sustained, focused intellectual work
− Cons
- Slower and more expensive than faster editor-native options
- Not the model I reach for first when the task is mostly visual iteration or quick tasks
GPT 5.4
GPT 5.4 is my primary general assistant. I reach for it for everyday conversations, general technical discussions, and building against modern tooling. It has fast, reliable tool calling and will often one-shot a feature or feedback pass when I give it a strong prompt, and the more up-to-date training data shows up in newer APIs, libraries, and patterns.
+ Pros
- My primary general assistant for everyday and technical conversations
- Fast, reliable tool calling with strong one-shot implementation ability
- Very strong for building modern software against newer tooling
− Cons
- Opus and Composer 2 still win for my default deep reasoning and in-editor loops
- I review output carefully on anything production-critical
Gemini 3.1 Pro
Gemini is my pick for fast, one-off chats and image generation in the app. It is not where I go for deep work, but it is great when I want a quick answer or need to generate an image without switching contexts. The full breadth of Google behind it makes it reliable for everyday general use.
+ Pros
- Great for fast, one-off chats where I want a quick answer
- My go-to for image generation in the app
- Strong multimodal and long-context workflows
− Cons
- Not where I lean for agent harnesses or agentic coding workloads right now — tool calling is not strong enough compared to my top picks
- Not my primary coding agent and less integrated into my editor-centric loop
Harnesses
Cursor
My main harness and clear number one. Cursor wins for me because I can use all the models I care about, Composer usage is generous, cloud agents are a huge plus for real, ad-hoc development work, and the overall UX, speed, and reliability are strong enough that it just works and just makes sense.
+ Pros
- Cloud agents are a major strength
- Best overall mix of model access, UX, speed, and reliability
- Multi-surface clients (agents UI, editor, CLI) that all consume the same models and agents on my subscription
− Cons
- Platform churn still happens from time to time, especially for research tasks
Claude Code
My second choice for coding harness. Claude Code is great for planning, writing, architecture reasoning, and long-form work tasks.
+ Pros
- Great for planning, writing, architecture reasoning, and long-form work tasks
- Great for structured outputs and critique
- New desktop app is great
- Can kick off work from mobile app
− Cons
- Not the most efficient for coding tasks compared to Cursor
- Not the most integrated with the editor
- Locked to Anthropic models only — no access to OpenAI, Gemini, or other providers
Tech stack
Languages
What I write most often — not the same thing as runtimes or frameworks below.
-
TypeScript · JavaScript · Go · Python
TypeScript and JavaScript for application code; Go for microservices and control planes; Python for tooling, notebooks, and ML-adjacent glue.
Runtimes
Where JS/TS executes — distinct from Hono or any web framework.
-
Bun · Node.js
Bun is my usual Node-compatible runtime for new JS services and scripts. I still reach for Node.js when tooling, libraries, or deployment targets expect it.
Web UI & client stack
Product-facing work: frameworks, styling, and build tooling.
-
React · TanStack Suite · Svelte / SvelteKit · Tailwind CSS · Vite · Astro
React with the TanStack Suite (Query, Router, and whatever else fits the feature); Svelte and SvelteKit when I want lean reactivity. Tailwind everywhere; Vite for builds; Astro for content sites like this one.
Backend & APIs
Services and HTTP boundaries — framework choice is separate from which runtime hosts it.
-
Go (Gin, Chi) · Hono · GraphQL federation · REST · WebSockets
Go for most service logic. Hono is my go-to small web framework for TypeScript HTTP APIs; I usually run it on Bun, but that is a hosting choice — not the same dimension as “what language” or “what framework.” GraphQL (including federation), REST, and WebSockets for APIs and realtime surfaces (dashboards, control planes, live updates).
Data, infrastructure, observability & ML
The resume-shaped layer: storage, delivery, signals, tests, and production ML patterns.
-
PostgreSQL · Redis · MongoDB
-
Kubernetes · Docker · Terraform · Helm
Multi-cluster and resilient deployments; GitOps and infra-as-code.
-
GitHub Actions
CI/CD pipelines and release automation.
-
Datadog · Prometheus · Grafana · OpenTelemetry
Metrics, traces, and dashboards aligned with on-call reality.
-
Playwright · Vitest · Jest · React Testing Library
-
LangChain · prompt engineering · model APIs
Plus evaluation and guardrail patterns for control planes in front of real models and traffic.