Slides

# Applied LLMs for Life Sciences

## 14-week course map

**21 contact hours · 7 guided hours · 28 hours total**

Weeks 1–14 count teaching sessions; numbering skips the holiday week.

---

# Cumulative skills, independent weekly cases

## Start with a tiny chat, then complete bounded voting-data tasks

<div class="workflow-row">
  <div class="workflow-step"><strong>Weeks 1–3</strong><br>First chat + build with an LLM</div>
  <div class="workflow-arrow">→</div>
  <div class="workflow-step"><strong>Weeks 4–6</strong><br>Tools → MCP → verification loops</div>
  <div class="workflow-arrow">→</div>
  <div class="workflow-step"><strong>Weeks 7–11</strong><br>Evidence, evaluation + review</div>
  <div class="workflow-arrow">→</div>
  <div class="workflow-step human-gate"><strong>Weeks 12–14</strong><br>Three project blocks</div>
</div>

Week 1 asks everyday questions in a PydanticAI browser chat without data or tools.
Independent cases through Week 11 develop reusable skills. Data-analysis labs
share one licensed dataset; Week 5 uses Connect Four and Week 7 adds official voting explanations. The final three weekly labs
build one bounded evaluation project step by step.

---

# Engineering spine · 8 levels of AI assistance

## The course targets reliable levels 2–6 before discussing greater autonomy

<div class="levels-track" aria-label="Eight levels of agentic engineering">
  <div class="levels-node"><span class="levels-dot">1</span><span class="levels-label">Tab complete</span></div>
  <div class="levels-node"><span class="levels-dot">2</span><span class="levels-label">Agent workspace</span></div>
  <div class="levels-node"><span class="levels-dot">3</span><span class="levels-label">Context</span></div>
  <div class="levels-node"><span class="levels-dot">4</span><span class="levels-label">Compound</span></div>
  <div class="levels-node"><span class="levels-dot">5</span><span class="levels-label">Tools / skills</span></div>
  <div class="levels-node"><span class="levels-dot">6</span><span class="levels-label">Harness</span></div>
  <div class="levels-node"><span class="levels-dot">7</span><span class="levels-label">Background</span></div>
  <div class="levels-node"><span class="levels-dot">8</span><span class="levels-label">Agent teams</span></div>
</div>

<div class="levels-overview-grid">
  <div class="levels-overview-card"><strong>Weeks 1–3</strong><span>First chat app, model limits, agent workspaces, and context.</span></div>
  <div class="levels-overview-card"><strong>Weeks 4–6</strong><span>Plotting agents, an MCP server, and feedback with stopping conditions.</span></div>
  <div class="levels-overview-card"><strong>Weeks 7–11</strong><span>Retrieval, code actions, richer MCP capabilities, evaluation, and review.</span></div>
  <div class="levels-overview-card"><strong>Weeks 12–14</strong><span>Evaluation, observability, human approval, and transfer.</span></div>
</div>

<p class="levels-source">Adapted from Bassim Eledath, <a href="https://www.bassimeledath.com/blog/levels-of-agentic-engineering">The 8 Levels of Agentic Engineering</a>. The weekly decks introduce these levels alongside specifications, types, tests, tools, logs, permissions, and review.</p>

---

# Weeks 1–3 · coding with LLMs

## Exploration becomes a verified change

| Week | Topic | Independent evidence |
| --- | --- | --- |
| [1](week_01.html) | First PydanticAI chat and LLM literacy | `agent.to_web()` chat, an everyday question, and a year/source follow-up |
| [2](week_02.html) | Vibe coding | Voting-data checker plus failure case |
| [3](week_03.html) | A tiny app from a short brief | A few bullets in `idea.md` and a NiceGUI initiative browser built from scratch |

---

# Course access across the weekly labs

## Use your course key for the agents in your code

Copy `.env.example` to `.env` and replace only `LLM_API_KEY`; no PPQ signup or
payment is required. The course uses `LLM_BASE_URL=https://api.ppq.ai` and
`LLM_MODEL=deepseek/deepseek-v4.1-flash` by default.

The optional model alternative is `openai/gpt-5.6-luna`. Save and restart the
running app after changing `.env`.

Coding assistants use your own subscription or free tier. The instructor
communicates the PPQ allowance and expiry for your application agents. Keep the
key out of prompts, screenshots, and submissions; local calculations remain
available when model access stops.

[Application setup and troubleshooting](setup.html) · [Coding-assistant access](coding_assistants.html)

---

# Weeks 4–6 · tools, MCP, and verification

## Build one useful capability in each session

| Week | Topic | Independent evidence |
| --- | --- | --- |
| [4](week_04.html) | Ask a dataset, build the plots | NiceGUI, up to N PydanticAI workers, checked plots, and saved explanations |
| [5](week_05.html) | Connect Four: models, decisions, and MCP | Play with different opponents, then tour MCP architecture, capabilities, transports and extensions |
| 6 (not yet released) | Design verification and correction loops | Evidence-based feedback, a repaired attempt, and an honest stopping outcome |

Each lab includes an independent reference. Students do not need a finished app from
the previous week. Typed fields and model configuration are introduced when the
working capability needs them; faster students take on extension challenges.

---

# Weeks 7–9 · evidence and richer capabilities

## Search, calculate, and publish traceable figures

| Week | Topic | Independent evidence |
| --- | --- | --- |
| 7 (not yet released) | Retrieval and agentic RAG | A short search/read baseline, a cited agent answer, and an insufficient-evidence case |
| 8 (not yet released) | Code Mode versus controlled tools | Generated code, deterministic reference, and a justified architecture choice |
| 9 (not yet released) | MCP queries, plots, and artifacts | A bounded query, labelled SVG, and retrievable artifact with provenance |

Week 7 combines manual retrieval and agent-directed search in one session.
Week 8 uses the supplied controlled-summary reference alongside Week 4’s plotting-tool pattern. Week 9 adds SQL, plotting,
and resources to the MCP connection pattern introduced with Connect Four in Week 5.

---

# Weeks 10–11 · challenge and review the output

## Make failure visible and approval enforceable

| Week | Topic | Independent evidence |
| --- | --- | --- |
| 10 (not yet released) | Evaluate an agent against evidence | Case-by-case results, failure categories, and a reproducible comparison |
| 11 (not yet released) | Review and approve an agent's output | An editable draft, version-bound approval, and a checked local export |

These are independent practical cases. They introduce the evaluation and review
behaviors that students will use in the final application. The
workflow-pattern reference (not yet released) is available from Week 6.

---

# Weeks 12–14 · three project blocks

## Each checkpoint extends the same Python report assistant

| Week | Project block | Checkpoint |
| --- | --- | --- |
| 12 (not yet released) | **1 · Reproducible evidence** | Validated data, deterministic results, **X ≥ 5** purposeful candidates, initial tests and saved provenance |
| 13 (not yet released) | **2 · Selection and draft** | Confirmed figure selection, order and notes, structured LLM draft from that evidence, Typst preview marked as a draft |
| 14 (not yet released) | **3 · Review and export** | Human review, Python-enforced approval of the exported version, final Typst/PDF and reproducible package |

These checkpoints are the weekly lab deliverables. Earlier cases, including
Week 7 RAG, Weeks 5 and 9 MCP, and Weeks 10–11 evaluation and review, provide preparation
and reusable artifacts without additional mandatory project milestones.

---

# Evaluation project

## A beautiful report must also be reproducible and reviewable

Build a Python assistant for one question on the shared voting CSV. Generate a
result table and a configurable set of at least five purposeful candidate
figures. Compare them, select one or more, order multiple figures, and confirm
notes stating what they show and a caveat. The structured LLM draft uses only
selected evidence; approved content compiles with Typst.

| Required evidence | Assessment weight |
| --- | ---: |
| Scientific question and analysis | 25% |
| Visual communication and Typst report | 20% |
| Human-in-the-loop workflow | 20% |
| Reproducibility and engineering quality | 15% |
| Evaluation and failure behavior | 10% |
| Provenance, transparency, and limitations | 10% |

Confirm the selection before drafting; explicitly approve the final version
before final export. Editing alone is not approval. Clearly marked draft PDFs
are permitted during review.

---

# Evaluation project · scope

## One analysis and one story are sufficient

- **Required:** Python application logic and analysis, evidence review, a
  versioned audit, reproducible Typst/PDF package and evaluation
- **Recommended UI:** NiceGUI; other frameworks are allowed, including a
  separate frontend with a Python backend
- **Bonus:** agent queries and plots from their computed results, general CSV
  exploration, AI card judging or notes, multiple stories, per-part review,
  project management and cost tracking

Bonuses are not needed for full marks. The Rust teacher application illustrates
behavior; students implement their calculations in Python. See the
project brief (not yet released) for the complete requirements and bonuses.

---

# One dataset, many weekly cases

## Keep the data familiar as the engineering task changes

<div class="course-grid">
  <div class="course-card">
    <strong>Data integrity</strong>
    <p>Schema, historical missing values, repeated identifiers, dates, and provenance.</p>
  </div>
  <div class="course-card">
    <strong>Analysis</strong>
    <p>Turnout, yes-vote shares, proposal kinds, distributions, and comparisons.</p>
  </div>
  <div class="course-card">
    <strong>Reliable applications</strong>
    <p>Typed outputs, controlled tools, citations, evaluations, and human review.</p>
  </div>
</div>

Analysis uses the unchanged `dataset/voting.csv`; Week 5 uses generated Connect Four
positions and Week 7 adds official voting explanations for retrieval. Cases through Week 11 are independent. Weeks 12–14
extend one application through three reviewable checkpoints.

<p class="source-note">Data: Vasco (vascos), <a href="https://www.kaggle.com/datasets/vascos/swiss-popular-voting">Swiss Popular Voting, Kaggle</a>, version 1 (24 November 2024), MIT license. The <a href="https://gitlab.idiap.ch/devel/biss/workshops/aiforlifescience/-/blob/main/dataset/README.md">data card</a> documents 692 national proposal records, 22 columns, units, missingness, and limitations.</p>