Technology radar

Standard IaC for all TD-managed services. TypeScript constructs, Cloudformation under the hood, shared construct library growing.

Standard AI coding assistant for TD. Skills and MCP servers are built for it. Recommended for any developer writing backend or infrastructure code.

Standard for internal tool UIs. Jotai replaces Redux for most use cases. Starter template available.

The standard web UI for Claude at MCS. Use it for ideation, drafting, and quick questions. Backed by Synapse Gateway — no setup required beyond an API key.

The practice of writing clear, structured prompts: role, context, task, format, constraints. High leverage with no infrastructure cost. Every developer should have a working mental model of this.

Proven for internal knowledge base queries. BM25 + cosine hybrid retrieval on pgvector. See TD-010 study for detailed recommendation.

Our standard for new Java Lambda services. Native-image builds, fast cold starts, CDK integration proven in production.

VS Code extension for AI-assisted coding, using the same Synapse Gateway endpoint as Claude Code. Good option if you prefer staying in the editor. Less mature than Claude Code for complex multi-file work.

Reusable slash-command extensions for Claude Code, shared via the skill catalog. Pattern is proving out — skill-catalog and guide-creator are in active use. Worth learning if you write Claude Code regularly.

Validated in the KPI dashboard. Real-time ingestion at low cost. Trial — not yet connected to production pipelines.

Jira MCP in active pilot. Pattern proven but catalogue is thin — only one server in production. Expand before declaring adopt.

Constraining model output to a strict JSON schema. Reliable for data extraction and tool-calling pipelines. Still requires careful schema design and fallback handling — not fire and forget.

Used for this catalog. Excellent for content-heavy internal sites. Trial — not yet proven for anything with auth or dynamic data.

Static full-text search for Astro/static sites. Assess for this catalog — straightforward integration, on the roadmap.

CSS-first config is cleaner than v3. Breaking change from v3. Assess for greenfield; don't migrate existing v3 projects yet.

Using models for automated code review, test generation, or release notes in pipelines. Early experiments show value but latency, cost, and reliability aren't yet pipeline-grade. No team recommendation.

Lower accuracy than RAG on our corpus AND higher latency at scale. Assess only for narrow, stable, latency-sensitive domains. See TD-010 study.

Orchestrating multiple AI agents that hand off tasks between each other. Promising for complex workflows but failure modes are hard to predict and debug. Watch before committing to production.

Superseded by Jotai for new internal tool UIs. No new projects should use Redux. Migrate opportunistically.

TD standardised on AWS CDK. No new TD-managed services should use Serverless Framework. Existing services: migrate at next major revision.

Running or fine-tuning models in-house. High operational cost, security surface, and maintenance burden. Synapse Gateway exists precisely to avoid this. Don't build what you can route through.