From Prompt to Private

Season four: the clinic gets its own AI, on its own server, because patient data cannot leave the building. Free open-weight models on Ollama, a 16 GB GPU (and an honest CPU-only minimum), knowledge in PostgreSQL with pgvector, a staff assistant and a patient assistant on the kiosk and the phone — built with AI across twelve parts, every number measured, every wrong answer kept in.

12 parts · ~71 min of reading, in order, nothing assumed.

Start with Part 1

📦 Same repo, season four — tags private-02 onward 🦙 Ollama, the local model runtime the season uses

From Prompt to Private · Part 1

From Prompt to Private: A Clinic's Own AI, on Its Own Server

Season four puts a private AI on the clinic's own server. Why patient data rules out the cloud, the first real run on a 16 GB GPU, and what this season builds. Sep 18, 2026 9 min read

From Prompt to Private · Part 2

Choosing the Model: Llama, Qwen and Gemma Sit the Clinic's Exam

Three free models sit a twenty-question clinic exam. Grounded answers, refusals, JSON, tool calls and an embedding check decide which one the clinic runs. Sep 18, 2026 7 min read

From Prompt to Private · Part 3

What Server Does a Clinic Need? Memory Arithmetic, Four Users and 120 Watts

The memory arithmetic behind local models, measured. CPU-only waits, GPU concurrency up to eight users, power draw, and a buy-or-rent table for a clinic. Sep 18, 2026 7 min read

From Prompt to Private · Part 4

Installing the Runtime on the Server: systemd, Loopback Only, and the Port You Never Open

Ollama on Ubuntu the way a clinic should run it. A systemd service on loopback, keep-alive and parallel settings, warm-up, and proof the port is closed. Sep 18, 2026 6 min read

From Prompt to Private · Part 5

Talking to It from .NET: OllamaSharp, Streaming, and a Chat Panel in the Staff Pages

ClinicLive gets its first assistant screen. IChatClient over OllamaSharp, a streaming Blazor chat panel, and an answer that proves why the next two parts exist. Sep 18, 2026 5 min read

From Prompt to Private · Part 6

Giving It Knowledge, Part One: Chunks, Embeddings and pgvector in the Clinic's Own Database

The clinic's documents become searchable vectors. A heading-aware chunker, nomic-embed-text, pgvector with an HNSW index, and ingest that skips the unchanged. Sep 18, 2026 5 min read

From Prompt to Private · Part 7

Giving It Knowledge, Part Two: Retrieval, Citations, and a Refusal the Model Cannot Skip

The assistant now answers from the clinic's own documents. Cosine search with an audience filter in SQL, a distance cut-off, and a guard for citations. Sep 18, 2026 5 min read

From Prompt to Private · Part 8

The Staff Assistant Asks the Queue: Tool Calling, and a Guard Against Tools Nobody Asked For

The assistant answers a live question through a function call. get_queue over Microsoft.Extensions.AI, a code-level guard, and a model that narrated in Thai. Sep 18, 2026 5 min read

From Prompt to Private · Part 9

The Patient Assistant: Public Documents Only, on the Kiosk and in the App

Patients get an assistant that sees only public documents. A rate-limited endpoint, a kiosk panel in giant type, the app screen, and a refusal that proves it. Sep 18, 2026 5 min read

From Prompt to Private · Part 10

Does It Actually Work? Fifty Questions, Three Models, and the Number the Season Was For

The private assistant sits a fifty-question exam, end to end. Three models scored on grounding, refusals, safety and traps, nothing invented by the winner. Sep 18, 2026 6 min read

From Prompt to Private · Part 11

Running It for Real: Updates, Backups, Logs, and What a Month Costs

The runbook for a clinic's own assistant. A warm-up timer, model and knowledge updates, what to back up, the logs, and 27.7 kWh a month against a token bill. Sep 18, 2026 5 min read

From Prompt to Private · Part 12

The Private Retro: What a Clinic's Own AI Really Costs, and What It Got Wrong

The season in one page. The gallery, the metered build against three earlier seasons, every mistake by who made it, three unproven things, and the verdict. Sep 18, 2026 6 min read