Six programmes, each ending in something you can run on Monday.

Applied LLM Foundations

6 weeks · evenings 19:00–22:00 · Saturday clinic · foundational · S$1,980

You assemble a small assistant with a prompt contract, a logged API wrapper, and an evaluation set of 80–150 labelled cases. The Monday artefact is a script that returns a structured answer and a score against last week’s baseline.

• Prompt as contract. Inputs, outputs, refusals and a version number in the file name.
• Evaluation set first. You write cases before you chase a clever phrasing.
• Logging and cost. Token counts and latency on a 50-item slice.
• Structured outputs. JSON schemas the next tool can consume.
• Regression after edits. A one-page score table for prompt versions 1 through n.

Automation Studio: Agents & Workflows

8 weeks · evenings 19:00–22:00 · Saturday clinic · intermediate · S$2,480

You build a short agent graph with tool calls, a human checkpoint, and a dead-letter queue. The shipped artefact is a workflow that can pause, retry, and show an operator why it stopped.

• Task graph on paper. Nodes, tools, and the human gate before any framework.
• Tool wrappers. One function per side effect, with timeouts you can read.
• State and queues. What is stored, what is replayed, what is dropped.
• Failure log. A table of the last 40 errors with the prompt version attached.
• Operator view. A minimal screen or notebook that shows current jobs.
• Handover. A runbook of 2 pages for the person covering Saturday.

Retrieval & Data Engineering

7 weeks · evenings 19:00–22:00 · Saturday clinic · intermediate · S$2,880

You index a corpus you control, usually 400–2,400 documents, and measure retrieval on a held-out question set. The artefact is a search path that returns passages with citations, plus the notes on chunk size that actually moved the score.

• Corpus hygiene. Deduping, licences, and what never leaves the laptop.
• Chunking trials. Three sizes scored on the same 40 questions.
• Embeddings and keyword mix. Hybrid retrieval with a recorded blend.
• Citations. Every answer points at a passage you can open.
• Drift checks. A monthly re-score recipe even if you only run it once in class.
• Access notes. Who may query the index after the cohort.

Model Evaluation & Guardrails

5 weeks · evenings 19:00–22:00 · Saturday clinic · intermediate · S$3,400

You write release criteria for a model-backed tool: gold set, refusal policy, logging, and a go/no-go sheet. The artefact is an evaluation pack a second person can rerun without you in the room.

• Gold and shadow sets. What is labelled, what is held back.
• Task metrics. Time, match rate, escalation rate. Model “accuracy” sits behind the task.
• Refusal and leakage. Prompts that must not answer, with tests.
• Logging minimum. Prompt version, model alias, latency, outcome.
• Release sheet. The five numbers a lead signs before a wider rollout.

AI for Operations Leads

4 weeks · evenings 19:00–22:00 · non-technical · S$3,600

You scope a workflow, write the evaluation questions, and leave with a brief a developer or vendor can price. Studio time uses spreadsheets, sample tickets and structured notes rather than a production repository.

• Workflow map. Volume, cycle time, and the 20% of cases that consume the week.
• Vendor questions. Data handling, eval access, and who owns the prompt.
• Reading a score. Baseline versus model, on a set you can explain.
• Pilot design. Two-week sandbox, 100 cases, named reviewer.
• Handover pack. Risks, owners, and the metric you will watch in month two.

Team Enablement Sprint

3 days · on your site or in Bras Basah · 8–16 people · private quote

A closed cohort for one team and one workflow. Day one freezes the evaluation set. Day two builds the thinnest path that scores. Day three writes the runbook and the list of items that will wait for a later programme. Quote depends on group size, location and whether sample data must be synthesised on site.

• Pre-sprint call. Data rules, laptop setup, and the metric the sponsor cares about.
• Shared repo or shared folders. One artefact the whole room can open.
• Role split. Operator, reviewer, builder. Each seat has a job.
• Exit brief. What shipped, what failed, what needs a 6-week line afterwards.