Pravaha

Ask your recordings. Watch the answer.

Turn hours of recorded lectures and talks into a library you can search, question and quote, where every answer is a playable clip of the moment it was said.

Live app · How to test · Cloudinary usage · Run locally · Docs

---

Hackathon project by Team Code Blooded - [hackindia-team:pixels-to-products-cloudinary-ai-hackathon-2026:code-blooded]

Hackathon	Problem statement	Track	Team
Pixels to Products: Cloudinary AI Hackathon 2026 (HackIndia × Cloudinary)	PS-03	Track 3: Your Media-Savvy Startup	Code Blooded

Contents

• The problem
• What Pravaha does
• Screenshots
• How Cloudinary powers it
• Business case
• Architecture
• How to test it
• Quick start
• Scripts
• Project structure
• Quality and testing
• Security
• Deployment
• Documentation
• Team
• License

The problem

Colleges, clubs and coaching institutes record hundreds of hours of lectures and talks. Almost nobody rewatches them, because you can't search a video, skim it, or quote it. The knowledge is recorded, then lost.

What Pravaha does

Watch	Upload a raw recording. Cloudinary transcribes it, chapters it and streams it adaptively. No editing. The transcript follows along as it plays, chapters are one tap away, and Ask this session answers questions about just that recording.
Find	Search every session for what was said and land on the exact second.
Ask	Ask a question in plain language. The answer comes only from your recordings, and every claim links to a playable clip of the moment it came from. Progress streams live while it works, and the answer suggests follow-up questions. If the library doesn't cover a question, Pravaha says so instead of guessing.
Answer Reels	The moments an answer cites, from different speakers and sessions, stitched into one labelled video.
Concept Map	/concepts lines up every concept your lectures teach and how many different teachers explain it. Open one and hear the same idea from each lecturer, back to back, as one Cloudinary-spliced video, until one explanation clicks.
Learning Paths	/learn: type a topic and get a 3–5 step course, ordered basics first, drawn from moments across the library and edited into one video.
Data saver	One switch (and automatic on Save-Data or 2G/3G) serves smaller, lower-quality Cloudinary renditions: thumbnails 76% lighter, clips and reels 38–41%, and a streaming ladder that tops out at 0.9 Mbps instead of 3.4. Built for India’s mobile data.
Weak spots	Quiz mistakes are remembered on your device and come back as one Cloudinary reel of exactly the explanations you missed, with the answer and the second to jump to. Get a question right and it leaves the list.
Study notes	Every session has a study sheet at /notes/[id]: summary, key concepts, chapters and the quiz with answers, each point linked to its timestamp. Copy it as Markdown, download it, or save it as a PDF.
Revision reel	On /saved, the latest saved moments from any sessions play as one labelled video: a revision reel for the night before an exam.
Voice	Tap the microphone in the Ask bar and say the question, in English or Hindi. The browser turns speech into text, so no audio reaches Pravaha.
Installable	Add Pravaha to a phone's home screen; it opens full-screen and shows a friendly page when offline.
Study Packs	Every session gets a summary, key concepts, a quiz whose explanations play the moment the teacher explains it, and a "Session in 60 seconds" highlight reel. Pravaha generates them automatically from the transcript.
Moments	One tap turns any cited clip into a vertical, AI-cropped, subtitled short for WhatsApp or Instagram, with its own share page and preview card. The short is just a Cloudinary URL; nothing is rendered.
Insights	Organizers see what learners ask, the knowledge gaps the library can't answer yet (what to record next), and which Moments get shared. Every report downloads as CSV. Visitors can explore a read-only demo of the Studio.
Embed	Organizers copy one <iframe> from the Studio and put Ask inside Moodle, Canvas or any course page, for the whole library or one session. Only that route can be framed. See docs/EMBED.md.
Try it	Anyone can upload a short video at /try, no account: Cloudinary keeps the first 60 seconds, transcribes and chapters it, and about 20 seconds later it has a Study Pack and an Ask this session box. Trials are private and deleted after 24 hours.

ChatGPT gives you text. Pravaha gives you the moment your professor said it.

Screenshots

Shown in dark mode; the light theme is one tap away in the header and follows the system setting by default.

Home	Watch: follow-along transcript, chapters, Ask this session

Studio (public read-only demo): Insights	Cloudinary under the hood

_{The demo library: six machine-learning lecture excerpts from four NPTEL courses.}

_{Ask on a phone. Every screen is built mobile-first.}

How Cloudinary powers it

Cloudinary isn't just storage here; the product depends on it.

Capability	Used for
Upload Widget + signed upload preset	Browser-to-Cloudinary video upload with real progress, no server proxy
auto_transcription	Word-timed transcript: the corpus for Find and Ask, plus subtitles
auto_chaptering	AI chapters on the player's seek bar
Cloudinary Video Player, HLS sp_hd_lean	Adaptive streaming (720p / 360p / 180p) that survives slow mobile data
e_preview,fl_getinfo, fl_sprite	The player's AI highlights graph and seek-bar previews
e_preview:duration_6	AI hover previews on library cards: the session's most interesting moments in 6 silent seconds
so_/eo_ + c_fill,ar_9:16,g_auto + timed l_text captions + f_auto,q_auto	Moments: trimmed, subject-tracked, subtitled vertical clips
g_auto thumbnails and poster frames	Content-aware library cards, results and social preview cards
q_auto:low, c_limit/w_ sizing, sp_sd streaming profile	Data saver: every video, reel, thumbnail and the HLS ladder re-requested lighter for slow or metered connections
l_video:…,fl_splice + timed l_text labels	Answer Reels and Session in 60 seconds: moments from one or more sessions stitched into one video
Webhooks (notification_url, signature-verified)	Upload → transcribed → ready with no polling
Incoming transformation eo_60 on a second signed preset	Try it uploads: Cloudinary keeps only the first 60 seconds, so a public upload can't cost more

What we built on top:

• turning transcripts into time-coded segments and indexing them
• library-wide retrieval and ranking (Postgres full-text search)
• grounded Ask, with citations checked on the server and refusal when the library doesn't cover a question
• the Moment and Reel URL composers
• the learner and organizer experience

Every session page and answer has a Cloudinary under the hood panel that lists the real URLs behind it, with each transformation explained in plain words and an Open button.

Full detail: docs/CLOUDINARY.md.

Business case

Track 3 asks for something you could pitch. Pravaha's pitch, in short (full version, with every number tagged actual, estimate or assumption: docs/BUSINESS.md):

• Who pays: coaching institutes first (their recorded classes are their product, and doubts cost teacher time), then college departments, then companies with recorded trainings.
• Why it wins: the answer is a clip of their own teacher saying it, never a chatbot's guess; it says "not covered" instead of bluffing, and that becomes a knowledge-gap report. Each institute's corpus and learner questions become a moat.
• Why Cloudinary's pricing fits: no transcoding farm, no GPUs. Transcripts, chapters, streams, crops, reels and share cards are generated on demand and cached, so cost follows what is watched, not what is stored. Hosting one hour of content is about 28.8 Cloudinary credits if every rendition is generated, and Ask costs a fraction of a cent.
• Pricing hypothesis (to validate, nobody has paid yet): a free pilot for one batch, then a one-time processing fee per content-hour plus a monthly fee per active learner.
• The binding constraint, shown live: Cloudinary credits. /status shows usage and warns at 80% and 95%.

Architecture

flowchart TB
  subgraph Org["Organizer"]
    O["Studio · upload · Insights"]
  end
  subgraph Learn["Learner · phone or laptop"]
    L["Ask · Find · Watch · Concepts · Learn · Saved · Notes"]
  end

  subgraph Cloudinary["Cloudinary: the media plane"]
    U["Signed upload preset"]
    AI["auto_transcription · auto_chaptering · hi-IN translate"]
    T["Transformations on demand: Moments · Reels · cards · previews · data saver"]
    S["HLS streaming · f_auto · q_auto"]
    X["Tags + context + Search API"]
  end

  subgraph App["Next.js 16 on Vercel: the knowledge plane"]
    W["Signed webhook"]
    R["Retrieval · question understanding"]
    G["Grounded answer + citation validation"]
    P["Study Packs · Learning Paths · Concept Map · Notes"]
    ST["/status · /api/health"]
  end

  DB[("Neon Postgres: segments · full-text index · packs · insights")]
  M["Gemini → Groq fallback chain · circuit breaker"]

  O -- video bytes, signed --> U
  U --> AI
  AI -- webhook --> W
  W -- time-coded segments --> DB
  W -- study pack --> P
  P -- tags and context --> X
  L -- question --> R
  R --> DB
  R --> G
  G <--> M
  G -- validated citations + clip URLs --> L
  T -- clips · reels · shorts --> L
  S -- adaptive video --> L
  ST -. checks .-> DB
  ST -. checks .-> M
  ST -. credits .-> Cloudinary

Loading

• One Next.js 16 app (App Router: frontend and API together), deployed on Vercel.
• Neon Postgres holds sessions, time-coded segments, the full-text index and insights.
• Cloudinary handles everything to do with media.
• Gemini handles the AI, with a model fallback chain ending in Groq (gpt-oss-120b) as a backup provider, and output validated against a Zod schema. If every model fails, Pravaha shows the most relevant clips instead of an error.

Diagrams and reasoning: docs/ARCHITECTURE.md · docs/TRD.md.

How to test it

Judges: /judges maps the submission requirements to the Cloudinary features, with live URLs and a 90-second test.

Open the live app at pravaha-cyan.vercel.app. Learners need no login.

1. On the home page, type a question or tap a suggestion. Try "Why should the learning rate decrease during training, and how do adaptive optimizers handle it?": live progress appears, then one answer citing two professors from two institutes, plus an Answer Reel. Or ask in Hindi: "ओवरफिटिंग क्या है?"

2. Hover or tap a citation number to preview the quote. Play a source card: the clip starts at the exact moment. Then try Watch the answer (the Answer Reel), Open full session or Share as Moment.

3. On a session page, open Study: the summary, key concepts, a quiz whose explanations play the moment, and "Session in 60 seconds".

4. Search a phrase. The results jump to the second, across sessions.

5. Ask something off-topic ("Who won the IPL?"). Pravaha says the library doesn't cover it instead of making something up.

6. Press / or Ctrl/⌘ K anywhere to jump to the Ask bar.

7. Open Try it and upload any short video with speech (up to 50 MB). Follow the live pipeline, then ask it "What is this video about?" on its page.

8. Open Studio: a read-only demo of the organizer side, with every session's pipeline output and live Insights (knowledge gaps, most asked questions, answers rated helpful). Uploading to the main library and publishing need the organizer passcode at /studio/sign-in.

9. Open Status: live database health, every AI model, and the Cloudinary credits the media runs on.

10. Tap the microphone in the Ask bar (Chrome or Safari) and ask by voice, or switch it to Hindi with the EN / हिं button.

Quick start

Prerequisites:

• Node.js 20+
• pnpm 9
• Free accounts for Cloudinary, Neon and Google AI Studio

git clone https://github.com/Monolithic-Dev/Pravaha.git
cd Pravaha
pnpm install
cp .env.example .env.local   # fill in the values, see SETUP.md
pnpm db:migrate
pnpm dev                     # http://localhost:3000

SETUP.md walks through every account from zero and explains every environment variable:

Variable	Purpose
NEXT_PUBLIC_CLOUDINARY_CLOUD_NAME, NEXT_PUBLIC_CLOUDINARY_API_KEY	Public Cloudinary identifiers for the player and Upload Widget
CLOUDINARY_API_SECRET, CLOUDINARY_UPLOAD_PRESET	Server-side signing and the signed upload preset
CLOUDINARY_TRIAL_PRESET, TRIALS_PER_DAY	Optional: the /try preset (default pravaha_trial) and the daily trial cap (default 5; 0 turns trials off)
STUDIO_DEMO	Optional: on (default) shows visitors a read-only Studio; off keeps it behind the passcode
DATABASE_URL	Neon Postgres connection string
GEMINI_API_KEY, GEMINI_MODELS	Gemini key and the ordered model fallback chain
GROQ_API_KEY	Optional backup AI provider, used when every Gemini model fails
ORGANIZER_PASSCODE, SESSION_SECRET	Studio sign-in and the signing key for its session cookie
APP_URL	Public base URL (webhooks, share links, preview cards)

Scripts

Command	What it does
pnpm dev	Start the dev server
pnpm build / pnpm start	Production build / serve it
pnpm lint · pnpm typecheck	ESLint · tsc --noEmit
pnpm test	Unit tests (Vitest)
APP_URL=… pnpm test:e2e	Playwright smoke tests against a deployed app, desktop and mobile
pnpm eval:ask	Evaluate Ask: grounding, citations and refusals (docs/AI_EVALUATION.md)
pnpm db:migrate	Apply database migrations
pnpm preset:trial	Create or update the 60-second trial upload preset used by /try

Project structure

src/
  app/                 Routes (App Router)
    page.tsx             Home: Ask bar, how it works, library
    search/              Ask + Find results
    watch/[id]/          Player, chapters, transcript, Study Pack
    m/[segmentId]/       Shareable Moment page with its own preview card
    studio/              Organizer: upload, publish, insights
    api/                 ask · search · lectures · upload-signature · webhooks · organizer · insights · events
  components/          UI (client components where interactivity needs it)
  lib/                 Server logic: retrieval, grounded answers, citation checks, Cloudinary URL composers,
                       ingest, auth, rate limits, env validation
tests/
  unit/                Vitest
  e2e/                 Playwright smoke tests
  eval/                Ask evaluation set
scripts/               Migrations, eval runner, local end-to-end
docs/                  Product, technical, security and process docs
openapi.yaml           API specification (OpenAPI 3.0)

Quality and testing

• CI runs lint, typecheck, unit tests and a production build on every pull request and every push to main.
• Unit tests cover the parts that decide what users see:
  • citation validation and refusal
  • follow-up question cleaning
  • transcript segmentation
  • the Moment and Reel URL composers
  • Study Pack validation, chapter parsing and upload-signing policy
  • webhook signatures and passcode checks
• End-to-end smoke tests run in Playwright on desktop and mobile against the live deployment:
  • home page
  • Find lands on the exact second
  • Ask answers with a playable citation
  • Ask refuses an off-topic question
  • unauthenticated organizer requests and unsigned webhooks are rejected
• The AI eval (docs/AI_EVALUATION.md) on production, Oct 2: 11/11 passed: 8/8 answerable questions cited the expected session (including one in Hindi), 3/3 off-topic questions refused, 0 fallbacks, and 31/32 citations (97%) supported their sentence on a manual check.
• Accessibility: a skip link, visible focus rings, keyboard shortcuts, and every animation disabled under prefers-reduced-motion.

Security

• Organizer routes need a passcode session: a signed, HttpOnly cookie.
• Uploads are signed, short-lived and size-capped.
• Cloudinary webhooks are signature-verified.
• Ask is rate-limited per IP before any paid call.
• Model output is schema-validated, and citations are checked against what was actually retrieved, which defends against prompt injection.
• No credentials are committed anywhere in this repository or its history.

Details: docs/SECURITY.md.

Deployment

The app deploys to Vercel: every push to main deploys production, and pull requests get preview deployments. Environment variables live in Vercel, with secrets marked Sensitive. Step-by-step: SETUP.md §7.

Documentation

Doc	What's in it
docs/VISION.md	Why this is a company: wedge, customer, growth loop, roadmap
docs/PRD.md	Problem, users, requirements, acceptance criteria
docs/CLOUDINARY.md	Every Cloudinary capability, and what breaks without it
docs/TRD.md · docs/ARCHITECTURE.md	Stack, reasoning, diagrams
docs/DATABASE.md · docs/API.md · openapi.yaml	Schema and endpoints
docs/EMBED.md	Putting Ask inside an LMS: snippet, sizing, framing policy
docs/AI_EVALUATION.md	Ask's retrieval, prompt, citation validation, fallback and eval
docs/UX_UI.md · docs/DESIGN_SYSTEM.md	Screens, interaction and motion, design tokens
docs/SECURITY.md · docs/COST.md	Trust boundaries and cost controls
docs/TESTING.md · docs/OBSERVABILITY.md	Test strategy, logs and events
docs/DECISIONS.md	Every decision with alternatives and trade-offs, including what we changed and why
docs/IMPLEMENTATION_PLAN.md · docs/phases/	The 3-day build, phase by phase
docs/GIT_WORKFLOW.md	Branch per feature, PR, merge rules
docs/DEMO.md · docs/SUBMISSION_CHECKLIST.md	Demo script and submission tracking

Demo library and credits

The live library is six 2.5–3 minute excerpts from NPTEL (IIT/IISc's National Programme on Technology Enhanced Learning) machine-learning courses, licensed CC BY-NC-SA. The excerpts are shared under the same licence, for non-commercial demonstration, with credit to:

Session	Course	Institute
Overfitting, Underfitting and Regularization · From AdaGrad to RMSProp	Deep Learning, Prof. Prabir Kumar Biswas	IIT Kharagpur
The Bias–Variance Trade-off · Learning Rate Decay · Gradient Descent Variants and Momentum	Machine Learning for Engineering and Science Applications	IIT Madras
Underfitting and Overfitting in Practice	Practical Machine Learning with TensorFlow	IIT Madras

Pravaha is built for an institution's own recordings; these public lectures stand in for them in the demo.

Team

Code Blooded: @Mahakisore7 · @im-rk

Built in three days for Pixels to Products: Cloudinary AI Hackathon 2026, organized by HackIndia × Cloudinary.

License

Apache-2.0. Pravaha is built on Cloudinary, Next.js, Neon and Google Gemini.