Week 8 · Day 40 of 40 · Build AI

Final release

You ship: WorkBot live, an architecture document, a recorded 10-minute demonstration and the numbers. Builds on: Day 22, Day 37, Day 38, Day 39.

By the end of the day you can

  • Release WorkBot: deployed, documented and measured
  • Show every capability the program trains running, with its evidence

Final Release

Features shipped today

  • WorkBot: Released with its architecture document, a recorded demonstration and its numbers

The last day is a release. Deploy the final version, write the architecture document, record the demonstration and collect the numbers.

Deliverables

  1. WorkBot, deployed by the Day 35 pipeline into an isolated environment, with the eval scoreboard, tracing and runbooks live
  2. An architecture document (3-5 pages): system diagram, agent design, data and integration layers, deployment, decisions and tradeoffs, what you would change
  3. A recorded 10-minute demonstration: a user on 2 channels, a document and a spreadsheet handled, an integration acting on a real system, a provider failover, a security control holding, and a change shipped through the pipeline during the recording
  4. The numbers: eval scores by provider, cost per conversation, p95 latency, deployment time from a clean account, PRs merged per hour in the Day 22 fleet run

Walkthrough

Release WorkBot. The merged commit goes through the pipeline into an isolated environment and passes every smoke gate there. The architecture document is brought up to the release, with its numbers and what comes next. A recording shows the product doing its work.

Part 1

Release through the pipeline

Study first

  • The release as a routine, not an event: the same images, the same Terraform, the same release script and gates as every release before it.

Build

  1. Build the merged commit's two images, create a clean cluster, apply the platform, and release into the isolated tenant's namespace with infra/release.sh.
  2. Read the gate's result before calling it released.

Done when

  • Every smoke gate passes in the isolated environment, and its health check names the merged commit.
Release
Released into Harbor Dental's isolated namespace: chat 0.96, extract 0.92, triage 0.88, security 1.00, and health at the merged commit.
Technical notes
  • The first attempt failed its gate. The triage cases name Fieldstone's teams, and this release was Harbor Dental's, whose teams are a front desk and billing, so the suite rejected every answer as an invalid team. The smoke suite now runs as the customer its eval sets were written for, with the release's other settings. Eval sets per customer are the next step.
  • With Harbor Dental's configuration, which puts chat on the large model first, the chat suite cost $0.25 for 20 answers against $0.013 on the small route. A customer's configuration changes the bill as well as the answers.

Part 2

The architecture document

Study first

  • What a reader of an architecture document needs: the system, the agents, the data, the deployment, the decisions with their costs, the numbers, and what would change.

Build

  1. Bring docs/architecture.md up to the release: deployment and operations from Days 34 to 39, the numbers with their sources, and what would change next.

Done when

  • Every number has a source in the repository, and every gap is named.
MeasureValue
Eval scores, Anthropic's small modelschat 0.92, extraction 0.92, triage 0.96, security 1.00
Eval scores, OpenAInot measured: no key
Cost per chat turn$0.0008 on the small route
p95 latency of a chat model call3.1 s small, 4.8 s large
Deployment from a clean cluster512 s with the smoke suite
API throughput on one 4-CPU node846 calls a second at 200 workers, no errors
Pull requests merged by the agent fleet20.6 an hour
Monthly cost at 1,000 usersabout $180 on the recommended platform

Part 3

The demonstration

Build

  1. A recording of the product at work, a caption per scene: a policy question answered with its source, a prompt injection refused, a scanned invoice read and filed, a spreadsheet question answered in SQL, the eval scoreboard, and the jobs page.
  2. Name what the recording does not show and where it is proven. Provider failover is the Day 36 game day. A second channel is Day 32's signed webhooks. A change shipped through the pipeline is the release above.

Done when

  • Each scene shows one thing the product does, and nothing is staged.
Demonstration
From the recording: the assistant refuses a refund it is not allowed to approve, citing the policy, and offers to send it to the person who is.

The recording, 72 seconds: demo.mp4.

Resources

  • All course materials from Weeks 1-8