What the work actually looks like

The problem

Two contract developers had built a booking prototype and then left. The code ran on one over-provisioned EC2 instance, credentials lived in a .env file on that server, and nobody could say which parts of the system stored patient data. The next clinic on their list required a security questionnaire completed before signing.

The constraint

The questionnaire used a control set the team had never worked against, there was no budget for a compliance consultant, and the prototype had to keep taking bookings for the two pilot clinics the entire time.

What we did

• Rebuilt the infrastructure as code so environments could be torn down and recreated, then recreated the live one without downtime.
• Moved patient records into a private-subnet Postgres instance with encryption at rest and daily snapshots, and cut the remaining services off from the public internet.
• Replaced the .env file with AWS Secrets Manager and rotated every key that had ever been on that server.
• Answered the questionnaire control by control, attaching the actual configuration as evidence rather than a written assurance.
• Deleted everything the prototype did not need, which turned out to be most of it.

What it cost

Discovery and audit

$850

Infrastructure and migration

$6,400

Security questionnaire support

$1,100

Ongoing retainer from month three

$1,450 / month

Stack

• AWS ECS Fargate
• RDS Postgres
• Terraform
• Next.js
• GitHub Actions

The problem

Dispatch staff were keeping a second, manual spreadsheet because the dashboard was too slow to use at the counter. Every page load ran four unindexed queries across a table that had grown past 40 million rows, plus a live third-party API call in the request path.

The constraint

Peak season was three weeks away, the schema was shared with an internal tool the team had not built, and no changes could be made to the data model itself.

What we did

• Instrumented the slowest path first and found that two of the four queries returned data the page never rendered.
• Added indexes on the columns the actual queries filtered by, in a maintenance window, using concurrent builds so the table stayed writable.
• Moved the third-party call out of the request path into a cached job refreshed every 90 seconds.
• Replaced per-row rendering of a 500-row table with server-side pagination and a covering index.
• Left a query-count budget in CI so the same pattern cannot creep back in unnoticed.

What it cost

Performance audit

$1,200

Query and caching work

$3,800

CI guardrails

$900

Stack

• PostgreSQL
• Redis
• Node
• Datadog
• GitHub Actions

The problem

The founders had a validated seller list and a conference booth booked, but nothing to show at it. They needed a site that could take seller applications and a small number of pre-orders, without looking like a template.

The constraint

The trade show date could not move, and the first version had to run on a budget that ruled out a full marketplace build with escrow and per-seller payouts.

What we did

• Scoped to a single flow that would survive contact with real users: browse two sample categories, apply to sell, place a pre-order.
• Cut per-seller payouts from v1 deliberately, and said so on the page rather than implying a capability that did not exist.
• Built the marketing site and the checkout as one deployable, so there was no second project to maintain.
• Set up Stripe, tax handling, and the transactional email before the design was finalised, because those were the parts with external lead time.
• Wrote the runbook for the founders before handover, and walked them through their own deploy.

What it cost

Scoping workshop

$600

Design and build

$4,900

Payments, tax, and email setup

$1,400

Handover and runbook

$500

Stack

• Next.js
• Stripe
• AWS S3 + CloudFront
• Postgres
• Resend