Wearly — 10-Minute Demo Script (May 14, 2026)¶
Audience: SEIS 666 instructor (Daniel Yarmoluk) + class. Goal: Show a working, reasoning agent — a working agent. One-line pitch: "Other styling apps tell you what to wear. Wearly tells you why, and lets you argue back."
Pre-demo checklist (run 5 minutes before)¶
| ✓ | Action |
|---|---|
| ☐ | Streamlit Cloud URL open in tab 1 (warmed up — cold start is ~30s) |
| ☐ | https://eatedalsf.github.io/styling-agent/ open in tab 2 |
| ☐ | Local repo open in tab 3 with python main.py --compare ready to run |
| ☐ | Mobile-viewport browser sized to 420px wide |
| ☐ | Confirm wear_history.json has seed entries (so "haven't worn in 35 days" fires) |
| ☐ | Backup downloaded in case Cloud restarts mid-demo |
At a glance¶
| # | Section | Time | The point |
|---|---|---|---|
| 1 | Opening hook | 0:45 | The problem is real and measurable |
| 2 | What makes this an agent | 0:45 | Track B framing, not "I used ChatGPT" |
| 3 | Before / After delta | 1:30 | The rubric's 20% line item — show the numbers |
| 4 | Live agent: calendar + weather + reasoning | 1:30 | Tools in action, citations live |
| 5 | Reject + regenerate (the agentic loop) | 1:00 | Argue with the system, watch it adapt |
| 6 | Wardrobe queries + compact KG export | 1:00 | Same data, two new surfaces |
| 7 | Intelligent Book + verified citations | 1:30 | "Structure beats prompts" — proved |
| 8 | Who would pay for this | 0:45 | Real-product framing |
| 9 | Closing | 0:25 | Land the message |
| Total | ~10:00 |
1. Opening hook (0:45)¶
Say:
"How long did it take you to decide what to wear this morning? For the average person it's about 15 minutes — and most of that decision happens without checking the weather, without checking the calendar, and with no memory of what you wore last week. Multiply 15 minutes × 5 days × 50 weeks. That's 62 hours a year — over a full work-week — on a problem you don't actually want to think about."
Pivot:
"What if your closet could reason about your day for you — and explain itself line by line?"
2. What makes this an agent (0:45)¶
Say:
"Wearly is a Track B project. Track B says: build a multi-step agent that uses tools, makes decisions, and completes a real workflow autonomously. Here's how Wearly hits every property of that definition:"
Show the 7-step workflow diagram (book/03 in tab 2). One breath each:
- Orchestrates 4 external tools: calendar, live Open-Meteo weather, wardrobe, color rules.
- Makes branching decisions: dress vs. separates, gym vs. work, outerwear vs. gap.
- Pursues a goal, not a response: a complete outfit appropriate to the day.
- Recovers from unexpected input: every tool has a fallback path.
- Has a closed feedback loop: reject + regenerate.
- Keeps persistent memory: wear history with a 0.4 freshness floor.
- Every output line ties back to a numbered rule.
"Wearly executes a plan, consults rules, and explains itself line by line."
3. Before / After — the delta (1:30)¶
Say: "The course rubric weights the before/after delta at 20%. Here's the measurable delta."
Open Streamlit → Before / After Demo screen.
| Metric | Before (manual / generic AI styling tool) | After (Wearly agent) |
|---|---|---|
| Time to decision | ~15 min | <30 sec |
| Reasoning lines exposed | 0 | 7–12 per outfit |
| Weather integrated | rarely | always (Open-Meteo) |
| Calendar integrated | never | always (next event read) |
| Audit trail | none | every line cites a rule slug |
| Verified citations behind the rules | 0 | 10 across 8 evidence categories |
| Reject + explain mode | not available | first-class loop |
| Wardrobe gap detection | "you should buy something" | "missing outerwear for a gym occasion → here's the specific suggestion type" |
"It's not just faster. It's a different shape. The before is one black-box paragraph. The after is a transparent reasoning trail with citations down to peer-reviewed papers."
Pivot: "Let me show you the after, live."
4. Live agent: calendar + weather + reasoning (1:30)¶
In the Streamlit app:
- Click Today → fill the new "Today's context" field with "feeling tired, comfort over polish". Note: this is the new feature borrowed from a classmate's project — Daniel praised this paradigm in our previous class.
- Click Plan today's outfit →.
- The result appears. Read out loud:
- The occasion card (pulled from calendar).
- The weather card (live Open-Meteo or fallback note).
- The first 4 reasoning lines, pointing at the bracketed citations (
[occasion-rules#R3],[weather-rules#R4],[wardrobe-filtering-rules#R3]). - Open the Reasoning graph expander. Drag a node.
"Every node here is in graph/schema.md. Every edge is a typed relation. Every reasoning line above traces to a numbered rule in the Skill package, which traces to a verified citation in the Evidence doc. Three clicks from output to academic source."
Land: If a "you haven't worn this in 35 days — bringing it back today" note fires (it should, because of the seeded wear history on item C001), point at it: "Wearly remembers."
5. Reject + regenerate (1:00)¶
Say: "This is the agentic loop."
Reject one item with the reason "too formal for today". Click Regenerate.
The new outfit appears. Point at:
- The new reasoning line beginning "Skipping '...' — you flagged it as: too formal for today. [wardrobe-filtering-rules#R5]"
- The new freshness or fit-alignment note that fired because the pool shifted.
"I argued with the system. It responded. It told me what changed and why — line by line, with citations."
6. Wardrobe queries + compact KG export (1:00)¶
Click into Wardrobe → Ask your wardrobe (new panel). Click Worn 30+ days ago.
"This is the same wardrobe data, surfaced as a queryable layer. Pattern borrowed from a classmate's project that lets you query the knowledge graph directly."
Back to the outfit. Click Export today's reasoning as a compact knowledge graph (new button). A JSON file downloads.
"One outfit, one portable graph. Same shape as our schema. The instructor mentioned in class he wanted to see sessions saved as compact KGs — here it is."
Open the JSON in a text editor for two seconds. Point at the entities and relations keys. Close.
7. Intelligent Book + verified citations (1:30)¶
Switch to tab 2 — the published Pages site.
Walk through three pages, ~15 seconds each:
- Skills → Skill package overview. "This conforms to the agentskills.io standard, Advanced tier per the meta-skills lecture — YAML frontmatter plus executable validator script."
- Evidence → Evidence & references. Scroll to §7. "Verified primary citations across all nine evidence categories. No fabrications. Springer 2022, ACM MM 2017, ECCV 2018, ISO 11664-4, EUR-Lex GDPR, Privacy by Design 2009, Tim Miller 2019. Every one is real."
- Learning Graph → Concept map (interactive). Switch to hierarchical layout. "Twenty-eight reader concepts, Bloom-taxonomy tier each, prerequisite edges between them. This is what makes the project an intelligent textbook, not just docs."
"Structure beats prompts. This is what the instructor has been saying all semester — and this is what it looks like fully wired."
8. Who would pay for this (0:45)¶
Open tab 3 / docs/business-model.md (or screenshot).
Three buyers, one sentence each:
- Consumer: $4–8 / month — direct subscription.
- Personal stylist: $20–40 / month — multi-client dashboard + the compact-KG export per client.
- Corporate wellness / DEI: license — the body-positive contract becomes a feature, not a constraint.
Land: "The moat isn't styling. It's the audit trail. Anyone can wrap GPT in a prompt. Almost nobody has a rule-cited reasoning trail with verified primary citations behind it."
9. Closing (0:25)¶
"Wearly is a Track B agent that exceeds every Track B requirement: nine tools where two were asked for, seven workflow steps with documented fallbacks, before/after delta you can measure, and a documented audit trail down to peer-reviewed citations. It was built with Claude Code. The code, the tests, the Intelligent Book, and the live app are all linked in the repo. Thank you."
Point to GitHub URL on the closing slide. Invite questions.
Likely questions — pre-baked answers¶
| Q | A |
|---|---|
| "Is this really agentic if the runtime is deterministic Python instead of an LLM loop?" | Yes — and deliberately so. The agent surface (Skills + tool orchestration + reject loop) is standard agentic shape; the evaluator is rule-based for inspectability. Book ch 4 "Why the reasoning core is inspectable rules" has the full answer. |
| "Why not collaborative filtering / a learned outfit model?" | Privacy (single-user scope today, no cross-user signal) and explainability (a learned model can't cite a numbered rule). Han 2017 and Vasileva 2018 are named in §3.2 as the alternatives we deliberately did not pursue. |
| "How are the colors chosen?" | Three skin-tone palettes in color_rules.json informed by warm/cool color-theory convention. Itten 1961, Munsell 1905, and ISO 11664-4 are background frameworks; Wearly does not compute CIE distances at runtime. |
| "Body shape feels reductive." | Agreed — and Wearly treats body_shape as a user-declared proportion preference, never an image-inferred classification. Hokka 2024 in Fashion Practice is cited in §3.4 for the inclusive-design framing. |
| "What if the weather API is down?" | Documented seasonal fallback in weather-rules.md R6. The outfit still ships; the trail says so. |
| "How did you build this?" | Entirely with Claude Code (Claude Opus 4.7). Documentation discipline (no invented citations, honest [to verify] markers on what remains) is enforced by a test suite. |
Presenter run sheet¶
| Step | Command / click |
|---|---|
| Pre-demo | Streamlit URL warmed, Pages site open, local repo ready |
| §3 | Open the Before / After Demo screen, scroll to the new delta table |
| §4 | Today → fill Today's context → Plan today's outfit → |
| §4 | Expand Reasoning graph |
| §5 | Reject one item with a reason → Regenerate |
| §6 | Wardrobe → Ask your wardrobe → Worn 30+ days ago |
| §6 | Back to outfit → Export today's reasoning as a compact knowledge graph |
| §7 | Switch to Pages tab — Skills overview → Evidence §7 → Learning Graph |
| §8 | docs/business-model.md |
| §9 | Closing slide with GitHub URL |
Total preparation summary¶
- 237 passing tests.
python -m unittest discover -s tests - MkDocs site builds clean. Auto-deploys on push.
- 10 verified citations in the evidence doc, no fabrications.
- All 8 Track B deliverables met or exceeded (see
docs/track-b-evaluation.md).