Full-cycle build and production migration of a standalone AI application

2026 / FDE — design, implementation, infrastructure, production migration, and first-line incident response, solo / Python · TypeScript · React · Vite · ECS Fargate · Kubernetes · Terraform · LLM API · RAG

design to production, incl. incident response

SCOPE — every layer, one person

6 days to production

DELIVERY — once unblocked

pytest 699 passed

TEST

The problem

On projects that put AI into real operations, there is a wide gap between “the PoC runs” and “the customer can actually use this.” Authentication and authorization, constraints on the customer’s side, and whether users can verify where an answer came from. Many projects stall at the PoC stage without crossing it.

Design decisions

Established the ability to show grounding — which document, and where in it — as a requirement question before implementation began. The gap between the granularity the existing platform handled and the granularity the customer expected was demonstrated by walking the codebase layer by layer, then reflected in the design
Stood up the PoC environment as an isolated stack so verification could run without affecting existing systems
Once the customer’s environment requirements were settled, rebuilt the container runtime, persistence, and routing, and moved to production
Wrote generated results back to the searchable dataset and proved the full path from registration to searchable
Implemented authentication, authorization, and CSRF protection to fit the requirements. Chose to fail fast rather than swallow errors
The environment had no operational logging, so log persistence was implemented before going to production

Lead time

Of the 16 days from start to production release, 10 were spent waiting for access to the customer environment to be granted. Once access came through, implementation, verification, and production release took 6 days.

Approach

Stood up the PoC environment as an isolated stack, separate from production, so verification could run without affecting existing systems
Once the customer's environment requirements were settled, rebuilt the container runtime, persistence, and routing, and moved to production
Wrote generated results back to the searchable dataset, and proved the path from registration to searchable as one continuous operation
Implemented authentication, authorization, and CSRF protection to fit the requirements rather than deferring to off-the-shelf mechanisms. Failures surface early rather than being swallowed
The environment had no operational logging, so log persistence was implemented before going to production

Next case study → Taking a research PoC into the product — knowledge graph feature shipped to production Ask about this project →