How to choose a tech stack in 2026

The five questions, in order

• 1. Hiring market — How many senior engineers in this stack are reachable in your metro or time zone within 60 days? This constraint outranks every technical preference, because a stack you cannot staff is a stack you cannot maintain.
• 2. Workload constraint — Does anything in the system have a hard requirement — sub-10ms latency, GPU inference, strict transactional integrity, offline mobile sync? Hard constraints eliminate options. Absent one, most mainstream stacks are equivalent.
• 3. Team fluency — What can your current team debug at 2am? A stack that is 20% better on paper and unfamiliar in practice is a net loss for at least two quarters.
• 4. Compliance and hosting — SOC 2, HIPAA, FedRAMP, and data-residency rules constrain where you deploy and which managed services you may use. Decide this before architecture, not after the audit.
• 5. Exit cost — How much would it cost to leave this vendor or framework in three years? Managed Postgres is cheap to leave. A proprietary workflow engine with your business logic inside it is not.

Our default stack, and why

For a new product with no hard constraint: TypeScript end to end, Postgres as the primary datastore, one managed cloud (AWS, Azure, or GCP — pick the one your buyer already has a contract with), Terraform for infrastructure, and GitHub Actions for CI.

The reasoning is boring on purpose. One language means engineers move between frontend and backend without context switching. Postgres does relational, JSON, full-text, and vector search well enough to postpone every other datastore. A single cloud keeps the operational surface small enough for a team of eight.

If AI is core to the product, add exactly one Python service for the model layer and keep everything else in TypeScript.

Where teams overspend

Kubernetes before you need it. Under roughly twenty services or forty engineers, managed containers cost far less in operational time and hire far more easily.

Microservices at the start. Modular monolith first, split when a specific team boundary or scaling constraint forces it — not because the architecture diagram looks better with boxes.

Four datastores in year one. Postgres plus one cache handles more than most teams believe. Every additional datastore adds a backup story, an access-control story, and a person who has to understand it.

Getting a second opinion before you commit

Stack decisions are cheap to make and expensive to reverse, which is exactly the shape of decision worth an outside review. We run paid two-week architecture reviews that end with a written recommendation, a risk register, and a hiring-market assessment for the recommended stack — usually a rounding error against the cost of finding out in month nine.