How we build production AI.
In short: our engineering depth is agentic delivery — we build with Claude Code and a library of custom skills, the agent-driven workflow that lets a small senior team ship and verify production AI fast.
How does NewGenApps build production AI?
We deliver with an agent-driven workflow, not a hand-off. A small senior team runs the build inside Claude Code, backed by a growing library of custom skills that encode our playbooks — scaffolding, retrieval, evals, and release checks. The method is the depth: it lets us move fast without trading away the review, testing, and independent verification behind our production AI, proven promise.
How do Claude Code and skills speed delivery?
Claude Code puts an agent in the codebase; our skills make it repeatable. Each skill captures a senior engineer's way of doing one thing well, so the same rigor ships on every project instead of living in one person's head.
- Skills as playbooks — RAG pipelines, agent tool-use, eval harnesses, and deploy checks, encoded once and reused.
- Small team, senior output — the people on the call are the people who ship; the agent handles the toil.
- Verified, not vibed — evals, red-teaming, and regression tests gate every release.
What is our AI engineering stack?
We stay vendor-neutral by design — the aim is that nothing on this list locks you in. These are the tools we currently reach for, and why; each is chosen so it can be swapped as the frontier moves. We mention products where it helps, not to endorse them.
| Layer | What we use | Why |
|---|---|---|
| Agentic delivery | Claude Code + custom skills | Encodes our senior playbooks so a small team ships and verifies fast |
| Model access | Amazon Bedrock | Model-flexible inference — swap the model without re-architecting |
| Train & MLOps | Amazon SageMaker | Fine-tuning, pipelines, and deployment in one place |
| Compute | GPU compute on EC2 (NVIDIA hardware) | Throughput for training and low-latency inference |
| Build & spec | Kiro — AWS spec-driven agentic IDE | Spec-first agentic builds that stay reviewable |
| Verification | Evals, red-teaming, regression tests | Proof a feature is production-ready |
Are we locked to one AI model?
No. Amazon Bedrock keeps the model layer replaceable, so we can route across models — including Anthropic's Claude — and change our mind as the frontier moves. We go deep on the tooling and the delivery method, not on any single vendor's model. That is what keeps clients off a bet that ages.
We build with these services and are fluent in them — we don't claim badges we don't hold. We're building toward Anthropic credentials, not an official partner, and we let shipped, verified work speak.
Book an AI working session