Add recorded presentation badge to README; untrack live presentation script

- README adds a 'Watch the Presentation' section between Watch the
Demo and Project Files. Orange badge linking to the recorded
10-minute MSML641 presentation on Drive.

- ppt/LIVE_PRESENTATION_SCRIPT.md untracked. ppt/*.md added to
.gitignore. The script served its purpose for the recording; the
recorded presentation itself is the deliverable now. Local copy
stays on disk if needed for reference.

.gitignore

@@ -1,7 +1,9 @@
-# Slide deck artifacts — kept out of the repo, hosted on Drive instead
-# (see README "Project Files" section for the folder link).
+# Slide deck artifacts and presentation drafting — kept out of the repo,
+# hosted on Drive instead (see README "Project Files" section for the
+# folder link).
 ppt/*.pptx
 ppt/*.pdf
+ppt/*.md
 
 # Python
 __pycache__/

README.md

@@ -46,6 +46,18 @@ short_description: Guarded RAG support navigator for UMD students
 
 <br>
 
+### 🎬 &nbsp; Watch the Presentation
+
+<div align="center">
+
+[![Watch the recorded MSML641 presentation](https://img.shields.io/badge/🎬%20Watch%20the%20presentation-10%20min%20recorded%20talk-fb923c?style=for-the-badge&logoColor=white)](https://drive.google.com/file/d/1PcHtjZwQix7aYJyslwp-aIuHj8Bs1C9f/view?usp=sharing)
+
+</div>
+
+> Two-presenter walkthrough of the problem, architecture, evaluation, and live demo — recorded for the MSML641 final submission.
+
+<br>
+
 ### 📁 &nbsp; Project Files
 
 <div align="center">

ppt/LIVE_PRESENTATION_SCRIPT.md

@@ -1,491 +0,0 @@
-# EmpathRAG — Live Presentation Script
-
-**Two-presenter, ~10-minute recorded PPT.** Open this file in **rendered Markdown view** while presenting (GitHub web view, VS Code preview, Obsidian, etc.) — the speaker blocks render as visual cards.
-
----
-
-## 🎯 How to use this script
-
-| Marker | Meaning |
-|---|---|
-| 🔷 **Mukul** | Mukul speaks the lines below |
-| 🔶 **Karthik** | Karthik speaks the lines below |
-| *(italics, plain text)* | Stage direction — do not say aloud |
-| Empty line **between blocks** | ~1-second silent pause / handoff |
-| Empty line **inside a block** | Brief breath, same speaker continues |
-
-Speakers are visually separated by **alternating card colors** (blue / orange) so you can spot at a glance who's about to talk. Each line within a card is one breath group — read one sentence, look up at camera if you want, then read the next.
-
----
-
-# 🎬 SLIDE 1 — Title
-
-`[ 0:00 – 0:12 ]`
-
-> 🔷 **MUKUL**
->
-> Hi everyone, I'm Mukul.
-
-> 🔶 **KARTHIK**
->
-> And I'm Karthik. Together we built EmpathRAG.
-
-> 🔷 **MUKUL**
->
-> It's a guarded conversational support navigator for UMD students.
-
-> 🔶 **KARTHIK**
->
-> Ten minutes, four sections — the problem, the architecture, the evidence, and a quick demo.
-
-> 🔷 **MUKUL**
->
-> Let's start with why we built it.
-
----
-
-# 🎬 SLIDE 2 — The Problem
-
-`[ 0:12 – 0:55 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> When students reach for a chatbot in distress, two things tend to go wrong.
->
-> The first one is fabrication.
->
-> The model invents a phone number, a service, an eligibility rule that doesn't exist.
->
-> The example on the right is exactly that — a number that *looks* official and isn't.
-
-> 🔶 **KARTHIK**
->
-> And the second is the opposite problem — missed signals.
->
-> Generic models tend to soften language that signals real distress.
->
-> They reassure when they should be intercepting.
->
-> And in a vulnerable moment, either failure is dangerous.
-
-> 🔷 **MUKUL**
->
-> So our goal was to fix both of those without losing the conversational quality students actually need.
-
----
-
-# 🎬 SLIDE 3 — The Headline Result
-
-`[ 0:55 – 1:30 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> This is the headline number — same Llama 3.3 70B model, two configurations.
->
-> On the left, the unguarded model misses escalation 9 times out of 28.
->
-> On the right, our guarded pipeline misses zero out of 28.
->
-> And the 95% confidence intervals don't overlap.
-
-> 🔶 **KARTHIK**
->
-> What's important here is that we didn't tune *for* this benchmark.
->
-> This is an external safety contract — does the system intercept crisis language, yes or no.
->
-> That's the bar a system like this should be evaluated against.
-
-> 🔷 **MUKUL**
->
-> Now — how do you actually get from a 32% miss rate to zero, with the same model underneath?
->
-> That's the architecture.
-
----
-
-# 🎬 SLIDE 4 — The Pattern: Plan and Rephrase
-
-`[ 1:30 – 2:05 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> The core pattern is what we call *plan and rephrase*.
->
-> We separate *what* the system says from *how* it says it.
->
-> A deterministic planner picks the route, the safety tier, and the resources.
->
-> Only then does the LLM rephrase the plan into natural language.
-
-> 🔶 **KARTHIK**
->
-> So the LLM never decides what to recommend. It only paraphrases.
->
-> Which means it physically cannot invent a resource the planner didn't already authorize.
->
-> That's the safety contract. And it's the foundation everything else builds on.
-
----
-
-# 🎬 SLIDE 5 — Architecture
-
-`[ 2:05 – 2:55 ]`  ·  *Click. Densest slide — pace ~10% slower.*
-
-> 🔷 **MUKUL**
->
-> Here's the full pipeline. Five layers.
->
-> Stage-1 is a lexical pre-check that catches crisis language before anything else runs.
->
-> If it passes, the router classifies the message into one of sixteen routes.
->
-> Then curated retrieval pulls from a verified UMD resource registry — no open web, no Reddit at inference.
-
-> 🔶 **KARTHIK**
->
-> From there the planner builds a deterministic response plan.
->
-> The rephraser turns it into natural language.
->
-> And the post-rephrase verifier rejects anything that drifted off the plan.
-
-> 🔷 **MUKUL**
->
-> And to be clear about what's actually trained here, because it matters:
->
-> The route classifier is TF-IDF plus logistic regression, trained on our 216-72-72 UMD dataset, about 86% test accuracy.
->
-> The LLM is Llama 3.3 70B via Groq — pretrained, *not* fine-tuned by us.
->
-> The templates and crisis regex are deterministic.
->
-> That separation between learned and hand-crafted is the entire safety story.
-
----
-
-# 🎬 SLIDE 6 — Iteration 1: Open Retrieval Baseline
-
-`[ 2:55 – 3:25 ]`  ·  *Click.*
-
-> 🔶 **KARTHIK**
->
-> The architecture you just saw is version three. Let me walk you through how we got there.
->
-> Our first version was a five-stage pipeline with open Reddit retrieval and single-turn responses.
->
-> And on standard metrics, it actually scored well.
-
-> 🔷 **MUKUL**
->
-> The problem only showed up when we ran adversarial probes against it.
->
-> Four specific failure cases told us the architecture had to change.
-
----
-
-# 🎬 SLIDE 7 — Why the Baseline Failed (Four Cases)
-
-`[ 3:25 – 4:10 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> Case one — generic emotional prompts dropped into a default route, with no specificity.
->
-> Case two — F-1 framing on turn one hijacked every later turn in the session, even after the topic moved on.
-
-> 🔶 **KARTHIK**
->
-> Case three was the worst one — a counselor was reported to have suggested harm.
->
-> And the system routed it to *academic-setback*, which implicitly validated the authority figure.
->
-> Case four — under explicit pressure to agree, the rephraser leaked a "you're right" capitulation before the verifier caught it.
-
-> 🔷 **MUKUL**
->
-> Each one of these became an architectural fix in the next iteration.
-
----
-
-# 🎬 SLIDE 8 — The Architectural Response
-
-`[ 4:10 – 4:35 ]`  ·  *Click.*
-
-> 🔶 **KARTHIK**
->
-> So we layered the response across four areas.
->
-> Listening for tone, instead of pushing resources immediately.
->
-> A trust boundary on the LLM.
->
-> Session-level state to carry context that should carry — and decay it when it shouldn't.
->
-> And a dedicated route for authority misconduct.
-
----
-
-# 🎬 SLIDE 9 — The Listening Layer
-
-`[ 4:35 – 5:15 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> Let's go deeper on the listening layer, because it's the most user-visible change.
->
-> Four stages — LISTEN, PERMISSION, OFFER, CLARIFY.
->
-> A student is *heard* before being routed.
-
-> 🔷 **MUKUL**
->
-> Turn one validates without dumping resources.
->
-> Turn two names a few options but asks permission before pushing further.
->
-> Turn three offers the full plan.
->
-> And CLARIFY catches single-word or incomplete replies, so the system doesn't barrel forward on insufficient input.
-
-> 🔶 **KARTHIK**
->
-> The practical effect is that the chatbot stops feeling like an FAQ bot.
->
-> And starts feeling like someone who's actually paying attention.
-
----
-
-# 🎬 SLIDE 10 — The Trust Boundary
-
-`[ 5:15 – 5:50 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> The other major change is the trust boundary on the LLM.
->
-> After the model rephrases, the verifier checks for drift.
->
-> Fabricated resources, scope creep, capitulation under pressure, AI-tells, length blow-up — any of those, and we fall back to the deterministic template.
-
-> 🔶 **KARTHIK**
->
-> And just to underline this — crisis content never enters the LLM path *at all.*
->
-> The Stage-1 pre-check intercepts it before the model ever sees the message.
-
----
-
-# 🎬 SLIDE 11 — Architectural Response (continued)
-
-`[ 5:50 – 6:15 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> And the session layer carries context where it actually should — F-1 status, prior offers.
->
-> Then it decays that context after two silent turns.
->
-> So the system can fully shift topic without one early signal dominating the rest of the conversation.
-
----
-
-# 🎬 SLIDE 12 — Datasets
-
-`[ 6:15 – 6:55 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> Karthik, walk us through the data.
-
-> 🔶 **KARTHIK**
->
-> Sure. We worked with three layers of data.
->
-> First, a curated UMD student-support conversational dataset — split 216, 72, 72 for training, dev, and test on single-turn routing.
->
-> Second, a 74-scenario multi-turn evaluation set for the safety contract.
->
-> And third, a verified registry of sixty-plus UMD resource URLs, each one annotated with a last-verified date.
-
-> 🔷 **MUKUL**
->
-> And Karthik led all of that — dataset curation, source verification, the annotation conventions for routing and safety tiers.
-
----
-
-# 🎬 SLIDE 13 — Per-Layer Ablation
-
-`[ 6:55 – 7:35 ]`  ·  *Click. This is the proof slide — land it.*
-
-> 🔷 **MUKUL**
->
-> So how do we know the layers are actually doing work, and not just sitting in the diagram?
->
-> We disabled each layer one at a time and re-ran the same 28-scenario escalation eval.
->
-> The Stage-1 lexical pre-check is load-bearing for missed escalation specifically.
->
-> Disabling it alone takes us from zero misses to twenty-two out of twenty-eight.
-
-> 🔶 **KARTHIK**
->
-> And the other layers protect orthogonal failure modes.
->
-> Registry filtering prevents fabrication.
->
-> The verifier catches LLM drift.
->
-> Each layer earns its place.
-
-> 🔷 **MUKUL**
->
-> This is also our cleanest answer to a fair question — *is the architecture real, or is it a scripted demo?*
->
-> If a layer weren't doing real work, disabling it wouldn't change the numbers.
->
-> The zero-to-twenty-two swing is the proof.
-
----
-
-# 🎬 SLIDE 14 — Targeted Failure-Mode Sweeps
-
-`[ 7:35 – 8:05 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> Beyond the headline eval, we ran five targeted sweeps to stress specific failure modes.
->
-> Rephraser drift across 29 cells.
->
-> F-1 stage and ISSS contract across 12 cells.
->
-> 25 sycophancy probes, 16 prompt-injection probes, 18 fairness paired prompts.
->
-> All clean within the stochastic LLM tolerance.
-
----
-
-# 🎬 SLIDE 15 — Honest Bounds on the Claims
-
-`[ 8:05 – 8:35 ]`  ·  *Click.*
-
-> 🔶 **KARTHIK**
->
-> A quick reality check before we move on.
->
-> N is twenty-eight escalation scenarios. That's small.
->
-> We're not claiming zero missed escalation in deployment.
->
-> What we *are* claiming is zero versus the unguarded baseline's nine, with non-overlapping confidence intervals.
-
-> 🔷 **MUKUL**
->
-> And the data is synthetic — real student phrasing is messier than anything we've evaluated on.
->
-> This is prototype-stage evidence, not a deployment claim. We say that out loud because it matters.
-
----
-
-# 🎬 SLIDE 16 — What's Next
-
-`[ 8:35 – 9:00 ]`  ·  *Click.*
-
-> 🔷 **MUKUL**
->
-> The roadmap follows the gaps we just admitted.
->
-> A real student dataset.
->
-> A RoBERTa router on top of the rule layer.
->
-> Scheduled URL re-verification.
->
-> And a multilingual opener for international students.
->
-> Now let's see the system in action.
-
----
-
-# 🎬 SLIDE 17 — Live Demo
-
-`[ 9:00 – 9:45 ]`  ·  *Click. Embedded 30-second auto-loop GIF starts playing on slide entry.*
-
-> 🔷 **MUKUL**
->
-> This is the system live. Let's narrate as it streams.
->
-> The student opens vaguely.
->
-> Notice — no resources dumped, just validation. That's the LISTEN stage doing its job.
-
-> 🔶 **KARTHIK**
->
-> Then they add context. Watch what the system does — it asks permission before offering anything.
-
-> 🔷 **MUKUL**
->
-> Permission granted. OFFER surfaces real cards, real links from the verified registry.
-
-> 🔶 **KARTHIK**
->
-> And when the student says "ok," the system doesn't re-render the same template. It advances.
->
-> The full five-minute demo with three more scenarios — F-1, substance use and confidentiality, crisis with sycophancy resistance — plus the Support Plan export, is linked in the README.
-
----
-
-# 🎬 SLIDE 18 — Thank You
-
-`[ 9:45 – 10:00 ]`
-
-> 🔷 **MUKUL**
->
-> That's EmpathRAG. Thanks for watching.
->
-> Code, datasets, evaluations, and the full write-up are all on GitHub.
-
-> 🔶 **KARTHIK**
->
-> Happy to take any questions.
-
----
-
-## ✅ Pre-record checklist
-
-- [ ] Both webcams placed in same corner. Don't move them per slide.
-- [ ] One full dry-run with a stopwatch. If you hit 10:30, cut the longest sentence on the slide where you went over.
-- [ ] Slide 5 (Architecture) and Slide 13 (Ablation) — pace yourself ~10% slower. Those are the dense slides.
-- [ ] Slide 17 — Mukul's first line should hit as the GIF's first response streams. Time the entry.
-- [ ] Don't say "uhh" — pause silently. Pauses read as confidence on camera. Filler words don't.
-
-## ✂️ If you go long — sentences to cut first
-
-| Slide | Cut this sentence |
-|---|---|
-| 4 | *"And it's the foundation everything else builds on."* |
-| 5 | *"The templates and crisis regex are deterministic."* |
-| 7 | *"Each one of these became an architectural fix in the next iteration."* |
-| 11 | *"So the system can fully shift topic without one early signal dominating the rest of the conversation."* |
-| 13 | *"Each layer earns its place."* |
-| 14 | Drop one sweep from the list (any of them) |
-| 15 | *"We say that out loud because it matters."* |
-
-## 💬 Q&A defense cheatsheet
-
-- **"Is this trained?"** → Slide 5 already named what's trained. Reference it.
-- **"Is this real?"** → Slide 13 ablation. Disabling Stage-1 changes the numbers, ergo the layers are doing real work.
-- **"What about substance / privacy / typos?"** → All have dedicated routes (`substance_use_concern`, `privacy_confidentiality`, typo-aware crisis detection). Demo'd in the linked five-minute video.
-- **"Why not fine-tune the LLM?"** → Plan-and-rephrase is the architectural commitment. Fine-tuning the LLM doesn't give you the deterministic safety contract.
-
-## 🔗 Continuity bridges (if you forget a transition mid-recording)
-
-| Going from → to | Bridge sentence |
-|---|---|
-| Slide 3 → 4 | *"How do you get from 32% to zero with the same model? That's the architecture."* |
-| Slide 5 → 6 | *"That's version three. Let me walk you through how we got there."* (Karthik takes over) |
-| Slide 7 → 8 | *"Each became an architectural fix."* (Karthik opens 8 with "So we layered the response...") |
-| Slide 11 → 12 | (Beat — Mukul hands to Karthik with "Karthik, walk us through the data.") |
-| Slide 12 → 13 | *"How do we know the layers are actually doing work?"* |
-| Slide 16 → 17 | *"Now let's see the system in action."* |