Technical deliveryat scale, and AIin production.

Suicide had taken people close to home, and each time nobody around them knew. Software does not fix that, but it can sit in the gap: one hour a week with a therapist, and nobody there for the other 167.

A model that forgets is not a companion. Continuity across months means solving retrieval, context and cost at once.

Treat the model as one component rather than the system, and design the layers around it: what is retrieved, what never reaches it, and what happens when it is wrong.

A 120-endpoint async FastAPI service, six providers behind capability tiers, hybrid retrieval over Firestore vector search.

Thousands of conversations across three platforms, on an architecture that came from psychiatry papers rather than product blogs.

AI assistants write framework-default code. On a bespoke platform that is wrong in ways which pass a glance and fail review.

Tiered repository access across four regions: a standard assuming full access solves it for the wrong half of the team.

Ground the assistant in the real codebase. Catalogue what exists so it never has to invent.

A versioned reference pack, and a four-stage build-and-review pipeline with a person at every checkpoint.

AI-assisted work is now reviewed, conformant and auditable, with or without core file access.

An automated attack on a live national campaign platform, mid-event, driving high volume through public forms.

Stopping it without stopping donations, where a false positive is a supporter who gives up halfway through giving money away.

Block at the edge, then treat the incident as evidence rather than noise to absorb and move on from.

Firewall rules at the edge, log evidence packaged for escalation, then a prioritised remediation plan.

The source was escalated to its host, who notified their customer. All four platforms brought to one standard.

A campaign wanted a points and rewards economy it could run itself: profile completion, first donation, team joining, fundraising milestones, daily activity, per-dollar accrual.

Webhooks retry and fire on state changes, so an award has to be safe to attempt repeatedly. And the front end cannot be trusted to say what somebody has earned.

Put the reward economy in the database. Twenty-five triggers as rows with a key, category, point value and enabled flag, all editable in admin.

A webhook receiver keyed by URL segment, an awarding engine with a three-mode frequency model, two custom tables including an append-only log, and a CMS module for the client.

A new trigger is a row in admin and a URL pasted into the CMS. No code change, no release, no engineer.

Four raffle transactions over five days recorded blank customer details. The money arrived, the entries were valid, and nobody knew who had paid.

One symptom, and nothing in the application code that could produce it. It was not reproducible from the platform alone.

Treat it as forensics. Correlate the records against raw server logs, request timing, user-agent strings and gateway responses, and rule the gateway out early.

Log-level reconstruction of real user journeys across specific browsers, wallets and devices, including a blank user-agent that proved which records a cron had written.

Three genuinely distinct root causes, each reproduced and explained rather than patched over.