docs/PITCH.md

Viveka — Pitch Scripts

Tight, on-brand pitches for mentor rounds and the final demo. Rubric anchors: innovation 40 / storytelling 30 / reward curves 20 / pipeline 10. Numbers in <ANGLE_BRACKETS> fill in once eval/holdout_eval.py lands real values.

60-second pitch (mentor rounds 1 & 2)

Viveka — Sanskrit for the wisdom to discriminate. We train LLM agents to predict whether an action is reversible before they take it, and emit a calibrated confidence on every prediction. Substrate: mocked Indian Digital Public Infrastructure — UPI for money, DigiLocker for documents, IRCTC for trains, with real NPCI/RBI error codes and business rules. Six-component reward, all deterministic state checks plus a Brier proper scoring rule on confidence — mathematically un-game-able, no LLM-as-judge anywhere. Two hero visuals: reward curve, base vs trained on the same axes, and a reliability diagram showing calibration improving alongside reward. Stack is TRL v1 GRPO + Unsloth 4-bit QLoRA on Qwen2-0.5B-Instruct, with 1.5B as a stretch run. By Sunday 5 PM you'll see a Hinglish agent that asks "are you sure?" before sending ₹50,000 to a flagged VPA — instead of guessing.

30-second hard pitch (mentor round 3 / demo opening)

Primary (when reward curve has landed):

Agents today guess on irreversible actions. Viveka teaches them to ask. Substrate: real Indian DPI — UPI, DigiLocker, IRCTC, with real error codes (UPI:5031 mandate cap, IRCTC:E2032 tatkal closed). Reward: six deterministic components, including a Brier score on confidence — Adarsh-paper anti-game-able. Reward curve: random <R>, frozen Qwen <Q>, Viveka <V>. Reliability ECE dropped from <X1> to <X2>. T4 adversarial safety: <S1>% → <S2>%. Empty competitive lane on GitHub.

Fallback (curve still rising at Round 3):

Viveka — wisdom to discriminate. Reward curve direction is clear and still climbing; let me show what's already locked: the reliability diagram. Same Qwen2-0.5B, base versus mid-training. ECE dropped from <X1> to <X2> on a sealed eval set. Calibration improving without LLM-as-judge — Adarsh's EMNLP 2025 anti-game-able territory. Six deterministic reward components, real Indian DPI error codes, Hinglish demo on the Space. Curve will land before submission; calibration story is already won.

---

90-second video — shot list

Voiceover at natural 1.0×, ~150 wpm. Burned-in captions for accessibility. Hook in first 3-5 sec. Total target 88 s (2 s buffer under 90 s cap). Output 1280×720, MP4 H.264, 6 Mbps + AAC 192 kbps.

Time	Screen	Voiceover	Capture	Hero overlay
0:00–0:05	HOOK split-screen: Hinglish prompt "Send rs5000 to mom for medicines" → red flash "FRAUD-WATCHLIST HIT, UPI:5050"	"AI agents guess on irreversible actions. Watch."	static slide → 1-frame red flash	red badge + UPI:5050
0:05–0:25	Terminal A (untrained Qwen2-0.5B): same prompt → executes blindly, predicted_reversibility=reversible (WRONG), confidence 0.91	"Today's frozen Qwen executes a UPI transfer to a flagged VPA. It mis-labels it reversible — at 91% confidence."	vhs .tape recording, zoom on confidence + reversibility	predicted_reversibility=reversible ✗ confidence=0.91
0:25–0:45	Terminal B (trained Viveka) + Gradio UI: same prompt → predicts irreversible, conf 0.88, emits confirm_with_user; UI shows "Are you sure? ₹5000 to vpa@flagged" in Hinglish	"Viveka pauses, predicts irreversibility, surfaces the watchlist hit, and asks the user — in Hinglish, the user's language."	screen-cap Gradio + terminal side-by-side	predicted_reversibility=irreversible ✓ confidence=0.88
0:45–1:05	Full-screen reward curve PNG (docs/reward_curve.png) with three lines + Ken-Burns pan-zoom	"Six deterministic reward components — including a Brier score on confidence. No LLM-as-judge. Random scores <R>%, frozen Qwen <Q>%, Viveka <V>%."	static PNG with motion + caption labels	random=<R>%, base=<Q>%, viveka=<V>%
1:05–1:20	Reliability diagram (docs/reliability.png) → cut to AQI delta bars (eval/aqi_delta.png) → cut to T4 adversarial table	"Calibration improves alongside reward. ECE drops from <X1> to <X2>. On adversarial scenarios — fraud watchlists, expired tatkal, revoked DigiLocker consent — safety jumps from <S1>% to <S2>%."	three quick cuts ~5 s each	ECE: <X1> → <X2>, T4 safety: <S1>% → <S2>%
1:20–1:30	Title card: "Viveka — wisdom to discriminate. UPI · DigiLocker · IRCTC. github.com/gowtham-sai-yadav/viveka-env" + QR code	"Viveka. The wisdom to discriminate. Repo, Space, model card linked below."	static end card; logo fade	tagline + URLs + QR

---

Recording protocol

Recording window: 14:00–16:00 IST on 2026-04-26. Submit at 16:00, deadline 17:00.

1. Tools (Day 1 evening): install vhs (brew install vhs) for terminal takes, OBS Studio 30.x for Gradio capture, iMovie for stitching. Avoid Loom (heavy compression).
2. Terminal theme (vhs .tape config): Set FontSize 22, Set FontFamily "JetBrains Mono", Set Theme "Catppuccin Frappe", Set Width 1280, Set Height 720, Set CursorBlink false, Set TypingSpeed 50ms, Set Framerate 60.
3. Pre-recording (13:30 IST): regenerate docs/reward_curve.png, docs/reliability.png, eval/aqi_delta.png with REAL data. Lock numbers into a recording_numbers.txt so the voiceover script has truth values.
4. Take 1 — untrained Qwen (docs/video/takes/01_untrained.tape): pipe Hinglish prompt to frozen Qwen2-0.5B; expect blind execute. Two takes, pick clean.
5. Take 2 — trained Viveka (docs/video/takes/02_viveka.tape): same prompt to trained checkpoint; expect confirm_with_user with VPA fraud reason surfaced.
6. Take 3 — Gradio UI (OBS, browser 1280×720, Chrome window-only, hide bookmarks bar): type Hinglish prompt, click submit, capture confirm modal. Mouseposé for click rings.
7. Take 4 — plot zoom-ins (Keynote/iMovie Ken Burns): full-screen PNGs with annotation arrows.
8. Voiceover: Debashis or Gowtham (whoever has clearer Indian English on the day); record in Voice Memos with WIRED earbuds (NOT MacBook built-in). Two takes per beat.
9. Captions: burn in via iMovie title overlay or Submagic auto-captions.
10. Stitch in iMovie: voiceover track 1, video track 2, plot stills track 3, captions track 4. Skip background music (90 s is too tight).
11. Click highlights: OBS Show Mouse Clicks filter or post-edit yellow circles in iMovie at click moments.
12. Export: MP4 H.264, 1280×720 @ 30 fps, 6 Mbps video, AAC 192 kbps.
13. Sanity playback: watch on phone (small-screen test) and on a 13" laptop. Bump font size if unreadable.
14. Upload by 15:30 IST as Unlisted YouTube; copy link into README + Devpost. Don't make Public — comment moderation distracts during judging.
15. Backup: commit the .mp4 to docs/video/viveka_demo.mp4 and push (Git LFS if >100 MB) so YouTube outage doesn't sink submission.

Skipped intentionally: background music, animated logo intros, slide decks. (Devpost guidelines + 90-s budget.)

---

Per-judge talking-point coverage (11 judges × 3 beats)

Beats: P = problem (0:00–0:25), E = evidence (0:25–1:20), A = ask (1:20–1:30).

Judge	Cluster	P	E	A
Sanyam Bhutani	Reproducibility / TorchForge	UPI watchlist hit shown via deterministic state	Reward curve, six components, no LLM-as-judge	Repo URL — Quick Start in 5 cmds
Yash Marathe	Reproducibility	Same	Reward curve overlay	Repo + HF Space + model card URLs
Adarsh Shirawalmath	Calibration / safety	Confidence 0.91 on a wrong call	Reliability diagram + AQI delta	"ECE <X1> → <X2>, T4 safety <S1>% → <S2>%"
Deepa Dhevannan	Calibration / safety	Same	Reliability + adversarial scenarios	T4 safety jump number
Ayush (Red Hat)	Production	UPI:5050 surfaced as real error code	Schema-strict Pydantic in terminal output	Production-realism close
Parshant (Red Hat)	Production	Same	Same	Same
Arkadip (Red Hat)	Production	Same	Same	Same
Aashay Sachdeva (Sarvam)	Indic	Hinglish prompt as hero example	Gradio UI Hinglish confirm flow	"Real DPI, real business rules"
Adithya Kolavi	Indic	Hinglish + ₹50,000 framing	UPI mandate cap, IRCTC tatkal references	Indic close
Nilesh Pandey	Indic	Same	Same	Same
Soumik Rakshit	Production / agentic	Frozen-vs-trained ask-loop contrast	confirm_with_user action live	Repo + Space + agentic-loop call-out

Every judge sees at least one beat from their cluster. Reliability diagram + Hinglish demo carry the heaviest cluster load (5 + 3 judges).

---

Risk register — 14:00-16:00 recording window

#	Risk	P	Impact	Mitigation
1	Training run not finished by 14:00	M	High	Pre-render template at 13:00 with <placeholders>; have 30-min plot-overlay swap path. Fallback voiceover script (above). Hard cut-off: if curve still flat at 14:30, pivot narrative to "calibration already locked" and feature reliability + AQI as hero.
2	Plots not generated / aqi_probe.py crashes	M	M	Run eval/holdout_eval.py + eval/reward_curve.py + eval/aqi_probe.py at 13:00 as smoke. If AQI fails, drop AQI bar and extend reliability shot to 8 s. Reward curve + reliability is minimum viable hero.
3	No quiet recording space at venue	H	H	Pre-book a breakout room for 14:00–14:30, OR record voiceover audio-only in a stairwell with Voice Memos + AirPods Pro mic; layer over screen captures in post. Test the spot at 13:30.
4	Mic quality bad	H	H	Use WIRED earbuds (3.5mm or USB-C), not Bluetooth (BT degrades to 8 kHz). Record 10-s test, listen on headphones BEFORE real takes. Submagic AI cleanup as fallback.
5	YouTube upload fails / processing stuck at 15:45	L	Critical	Upload by 15:30 as Unlisted at 720p. Backup: commit .mp4 to docs/video/ + Git LFS; link both URLs in README. Devpost accepts a direct repo-hosted MP4 if YouTube fails.
6	Video drifts > 92 s	M	M	Timecode-check at every edit pass. Target 88 s, hard cap 92 s.

---

Per-judge framing notes (preserved from earlier passes)

• Sanyam Bhutani: lead with verifiable rewards + clean PyTorch-native GRPO loop. Drop "deterministic state checks, no LLM-as-judge" — TorchForge agenda.
• Adarsh Shirawalmath: lead with proper-scoring-rule on confidence + reliability diagram. His EMNLP 2025 AQI paper is literally about catching mis-calibrated alignment.
• Aashay Sachdeva (Sarvam): lead with RLVR + Indic substrate. Don't overclaim "sovereign" — frame as "real DPI APIs, real business rules."
• Adithya Kolavi / Nilesh Pandey / Deepa Dhevannan: lead with Indian context — UPI ₹1L mandate cap, IRCTC tatkal windows, DigiLocker consent token expiry. Real, not toy.
• Red Hat trio (Ayush / Parshant / Arkadip): lead with production-systems realism — error codes, schema-strict Pydantic, fault tolerance in env step loop.
• Soumik Rakshit: lead with agentic asking-loop and trained-vs-frozen JSON-failure rate exposure (the methodology-honesty story).