AI & LLM Tooling
Want to propose a change, disagree with a placement, or add an interesting new technology? Please reach out to @Damian Szafranek. This Tech Radar is a living document - we keep it useful by updating it together as our projects, tools, and experiences evolve.
Scope
This section covers the tools, frameworks, and patterns we use to apply Large Language Models (LLMs) in both product development and developer workflows. It includes model access and routing (gateways, vendor abstraction, cost/limit controls), LLM pipeline and context orchestration (prompting, retrieval/RAG-style assembly, message construction), and AI-assisted development tooling (coding assistants and agentic coding approaches), along with guidance on where these tools provide clear value versus added complexity.
Overview
TRIAL
Promising technologies to explore and evaluate for adoption.
ASSESS
Interesting technologies requiring investigation.
HOLD
Not recommended for new work; plan migration from existing uses.
Technologies Breakdown
OpenRouter ADOPT
A practical “LLM gateway” that lets us access and switch between multiple model providers through a single API. Helpful for avoiding lock-in and choosing the best capability-to-cost model per use case.
OpenRouter simplifies multi-model usage: one integration can route to different vendors/models, making it easier to experiment, optimize costs, and respond to model availability or quality changes without rewriting our stack. It’s particularly useful for teams that want flexibility (try several models, compare results, tune budgets/limits) while keeping implementation effort low. Recommended as the default integration layer when a project expects to iterate on model choice or wants to reduce dependency on a single provider.
t3.chat ADOPT
The recommended AI chat assistant for augmenting development tasks like code explanation, debugging assistance, generating code snippets, writing documentation, and research. Leverage its capabilities while adhering to responsible use guidelines (verification, privacy).
Provides an exceptionally powerful AI chat interface (powered by models like Gemini 2.5 Pro) tailored for development workflows, offering strong capabilities in code understanding and generation. Considered the most powerful and cost-effective option currently available, priced very competitively at approximately $8. Adopted as the standard AI assistant to enhance developer productivity and knowledge acquisition due to its high capability-to-cost ratio.
v0 ADOPT
Adopted tool for rapidly generating initial UI components and prototypes from text prompts or images. Use the generated React code (based on shadcn/ui and Tailwind CSS) as a starting point for further refinement.
Proven to significantly accelerate the initial UI development and prototyping phase. Generates code based on our adopted UI stack (shadcn/ui, Tailwind), making integration smoother. Adopted as a valuable tool for quickly translating design ideas into functional code skeletons.
Haystack TRIAL
A framework for building LLM-powered pipelines (e.g., retrieval + prompting + post-processing) that helps structure “AI features” especially where context building and orchestration matter.
Haystack supports composing repeatable pipelines for use cases like internal knowledge assistants, document-centric Q&A, or other “LLM with context” flows. It’s on the radar because it works well for basic scenarios and provides a clear structure for connecting components (retrievers, prompt builders, model calls, evaluators). We keep it in evaluation mode to compare it with alternatives and to validate production concerns (observability, maintainability, integration patterns). Recommended when you need consistent pipelines, context assembly, and predictable behavior.
Agentic Coding Tools (ClaudeCode, OpenCode, Antigravity) TRIAL
Agentic coding tools that go beyond “chat + autocomplete” by running structured, multi-step workflows (plans, tasks, tool usage) to help implement features, refactors, migrations, and repetitive engineering work more efficiently.
These tools introduce an “agent” layer: you define goals or skills (and sometimes constraints), and the tool can iteratively propose changes, apply edits, run checks, and refine results. We’re trialing multiple options (Claude Code, OpenCode, Google Antigravity) to identify what best fits our workflows, security expectations, and day-to-day ergonomics. Use them for high-leverage tasks (scaffolding, refactoring, test generation, repetitive changes), but keep human ownership: verify outputs, review diffs carefully, and follow privacy rules (no sensitive data).
GitHub Copilot TRIAL
Continue exploring as an AI pair programmer integrated into IDEs for code completion and suggestion. Evaluate its specific strengths alongside for different tasks (e.g., inline completion vs. conversational interaction). Requires license management.
Offers powerful AI-driven inline code assistance. Trialing continues to assess its specific productivity benefits in comparison and conjunction with a chat assistant, and to refine best practices for leveraging both tools effectively.
Cursor TRIAL
Explore this AI-first code editor as a potential alternative or supplement to existing IDEs (like VS Code). Evaluate its integrated AI features, performance, compatibility, and overall developer experience on specific projects or by interested team members.
Represents a different paradigm with deeply integrated AI features. Trialing to understand if this approach offers tangible benefits over using VS Code combined with our adopted/trialed AI assistants, and to assess its maturity and suitability for our workflows.