Which AI model should you use — Claude, GPT, Gemini, Perplexity or Grok?

It depends on the task — and the real answer is usually more than one. I pick and route across frontier models by strength, with evals and fallbacks, so you get the best result without lock-in.

Which AI model should you use — Claude, GPT, Gemini, Perplexity or Grok?

It depends on the task — and the real answer is usually more than one. I pick and route across frontier models by strength, with evals and fallbacks, so you get the best result without lock-in.

AI consulting & engineering

Choosing and combining frontier AI models

Q: How do you actually implement AI in a business?

I find where AI creates real value, design the architecture, and build it end-to-end — running systems, not slideware.

Q: Can you build AI into a healthcare product?

Yes — I have architected clinical decision support, AI meal analysis and LLM health assistants on a custom model + vector database, with the privacy and reliability healthcare demands.

Q: Can you build a production mobile app in React Native?

Yes — I ship native iOS and Android apps from one Expo/React Native codebase, with payments, provisioning, offline-first flows and clean native integrations.

Q: How do you scale a Node.js backend?

With the right architecture: services split by load, event streaming, queues and caching, observability, and databases chosen per workload — proven on a platform that grew to 3M+ users.

Q: Can you build an MVP and then scale it?

Yes — I build MVPs end-to-end fast, then evolve the same codebase to production scale instead of throwing it away. Several of mine grew to acquisition or millions of users.

Q: How do you build a project with agentic AI development?

I drive AI coding agents (Claude Code, Codex, Cursor) with custom rules, hooks and context, spec-first — then gate everything with tests and AI review so speed never costs quality.

Q: How do you use CodeRabbit for code review and quality?

CodeRabbit reviews every pull request automatically — catching bugs, security issues and smells — as a gate alongside human review and a real test suite, so nothing weak reaches main.

Q: Can you build a custom AI model for my product?

Yes — self-hosted models on Ollama (Llama, Mistral, DeepSeek), fine-tuned or RAG-grounded, when you need privacy, control or a task an off-the-shelf API can’t do well.

Q: How do you build a RAG system?

Chunk and embed your data into a vector database, retrieve and rerank the right context per query, ground the model in it, and evaluate — so answers are accurate and cite your sources, not hallucinations.

There is no single best model — there is the right model for a task, and the discipline to route between them. I have shipped products on OpenAI, Claude, Gemini, Grok and DeepSeek together, so I choose by evidence, not hype, and design so switching providers is a config change, not a rewrite.

Pick by strength, not by brand

Claude is my default for coding and careful reasoning; GPT/Codex for breadth and tool use; Gemini for long-context and multimodal; Perplexity and Grok for live web research and fresh answers; self-hosted Llama/Mistral/DeepSeek when data can’t leave. On Leadtech’s aichatapp.ai I put GPT-4o, Gemini, Claude, Grok and DeepSeek behind one product — so I know how each behaves under real load.

Route, evaluate, and keep a fallback

The winning setup is rarely one model. I route each request to the model that does that job best, measure quality with evaluation sets, and keep a fallback provider so a single outage or price change never takes the feature down. That is how OptAI’s pipelines stay reliable across OpenAI and Cloudflare AI.

No lock-in by design

I abstract the model layer behind a clean interface (LangChain or a thin in-house router) so swapping or adding a provider is a config change. As the frontier moves — and it moves monthly — your product upgrades without a rebuild.

Key points

Match the model to the task: coding, reasoning, long-context, live research or private.
Route across models, evaluate quality, and keep provider fallbacks.
Abstract the model layer so you never get locked into one vendor.

Tools I reach for

Claude AIOpenAICodexGeminiPerplexityGrokOllamaLangChain

Have something like this in mind?

Get in touch

Choosing and combining frontier AI models

Pick by strength, not by brand

Route, evaluate, and keep a fallback

No lock-in by design

Related work

Have something like this in mind?

More questions I'm asked