DEV Community: Kaustubh Upadhyay

CineMuse 🎬 — an AI film mentor for the short films we can't stop making

Kaustubh Upadhyay — Mon, 13 Jul 2026 04:56:35 +0000

This is a submission for Weekend Challenge: Passion Edition

What I Built

Everyone I know who makes short films has the same problem: nobody watches their drafts.

Friends say "nice!" after three seconds. Feedback threads die. So the film you've poured your nights into gets edited on instinct alone, and you publish it still wondering — does the opening drag? Does the ending land?

CineMuse is a private screening room for those films. You upload your short film (animated or live action, up to 12 minutes), tell it your plot and what you meant to say, and Google Gemini watches the whole thing — every frame. Then it hands you:

🎯 The annotated cut — time-coded notes pinned as clickable markers on your timeline, Frame.io-style, except the AI wrote them. Click a marker, the film seeks there.
🎬 A scene-by-scene breakdown with real thumbnails (captured in your browser)
📈 An emotional arc graph of your story's beats — it draws itself as you scroll
🩺 A plot doctor — stakes, alternative directions, ending ideas, what to cut
💬 "Ask the Muse" — a chat grounded in what the AI actually watched. Ask "how do I fix my opening?" and it answers with timestamps from your footage.

No login. Sessions save in your browser. And because it's a passion project about passion projects, the whole UI is a cinema: velvet curtains that part when your film lands on the silver screen, a NOW SHOWING marquee with chasing bulbs, a film-leader countdown while the muse watches, and film strips rolling through the footer.

Demo

🔗 Try it live: https://cinemuse-orcin.vercel.app

Code

🗂️ GitHub: https://github.com/KaustubhUp025/cinemuse (MIT)

You can run your own CineMuse in about 3 minutes — all it needs is Node 18+ and a free Gemini key (no credit card):

git clone https://github.com/KaustubhUp025/cinemuse.git && cd cinemuse
cp .env.example .env.local   # paste your key from aistudio.google.com/apikey
npm run dev                  # → http://localhost:3000

There's deliberately nothing to install: no framework, no build step, zero npm dependencies. The frontend is vanilla HTML/CSS/JS (every animation hand-rolled), the backend is four small Vercel serverless functions, and the local dev server is a single Node file that mimics Vercel's runtime.

How I Built It

The interesting problem: how do you analyze a whole video on a free tier?

Video tokens are expensive, so CineMuse runs a map-reduce pipeline over the film:

The browser uploads the film directly to Gemini's Files API (a resumable, CORS-enabled upload — my server only signs the session, so the API key never leaves the backend and there are no proxy size limits).
MAP: gemini-flash-lite watches the film in ~20-second clips using videoMetadata time offsets at low media resolution — cheap, parallel, timestamped observations.
REDUCE: gemini-flash never sees a single video token. It gets the text observations + your plot + your focus areas, and writes the full critique as structured JSON (responseSchema) — scores, annotations, scenes, arc, plot doctor.
"Ask the Muse" reuses the saved observations as chat context — follow-up questions cost zero video tokens.

A 60-second film ≈ 7 API calls. The whole thing is vanilla JS + Vercel serverless functions — no framework, no build step, all the motion is hand-rolled CSS.

Fun bugs from the weekend: Google's resumable uploads demand 8MB chunk granularity (which killed my proxy plan — direct browser upload saved it), and Gemini's thinking tokens silently ate my chat token budget, truncating the muse mid-sentence.

Prize Categories

Cinemuse is built with extensive use of Gemini AI and that is why I would like to submit this project under Best Use of Google AI

The Passion Angle

CineMuse exists so that work finally gets seen properly: watched all the way through, taken seriously, and answered with the kind of specific, warm, timestamped feedback a mentor would give.

Built solo over the weekend. If you've got a draft sitting in a folder — give it a screening. 🎞️

Stack: Google Gemini (Files API + multimodal video understanding) · Vercel serverless · vanilla HTML/CSS/JS

MeetChecker : Automated meeting preparation using RunnerH

Kaustubh Upadhyay — Sun, 06 Jul 2025 12:21:48 +0000

This is a submission for the Runner H "AI Agent Prompting" Challenge

What I Built

I automated weekly meeting preparation using Runner H, turning a manual, repetitive workflow into a streamlined autonomous process. Runner H fetches upcoming meetings from Google Calendar, generates preparation documents with relevant summaries and pre-reading suggestions, creates reminders, and updates a tracker in Google Sheets—all automatically. This workflow ensures I am fully prepared for meetings with zero manual prep work, saving hours each week.

Demo

RunnerH chat link:- https://runner.hcompany.ai/chat/3dcc0e46-1a8f-483b-8cd4-a45fd2bb1c35

Demo-Link :- [https://youtu.be/7lqovSxiEZc]

How I Used Runner H

The prompt I used with Runner H is given as: -

You are MeetChecker, working with RunnerH, an autonomous agent connected to my Google Calendar, Google Docs, Google Drive, and Google Sheets.

Your task is to help me prepare efficiently for the upcoming week by:

1️⃣ **Access Google Calendar**  
- Fetch all meetings and events scheduled for the upcoming 7 days.  
- For each meeting, extract:
    - Title of the meeting/event
    - Date and time
    - Location (if any)
    - Attendees (if available)
    - Meeting description and agenda

2️⃣ **Generate Preparation Documents**  
- For each meeting, create a Google Doc titled:
  "Preparation - [Meeting Title] - [Date]"
- The doc should include:
    - Event details (title, date, time, location, attendees, description)
    - A summarized list of key points to prepare, based on the agenda
    - Potential questions I should be ready to answer or ask
    - A "Notes" section for me to fill during or after the meeting

3️⃣ **Pre-Reading Suggestions**  
- For each meeting, based on the title and agenda, search the web to find:
    - 3–5 relevant articles, papers, or resources to read for preparation
    - Summarize each in 2–3 bullet points
    - Provide direct links

4️⃣ **Create Reminders**  
- Schedule reminders in Google Calendar:
    - One reminder 24 hours before the meeting titled: "Prep Reminder: [Meeting Title]"
    - One reminder 1 hour before the meeting titled: "Join: [Meeting Title]"

5️⃣ **Weekly Overview Doc**  
- Create a Google Doc titled "Weekly Meeting Preparation - [Week Starting Date]"
- Summarize:
    - A table listing each meeting with date, time, and preparation status
    - Quick links to the individual preparation docs for each meeting
    - A prioritized list of meetings requiring in-depth preparation

6️⃣ **Optional: Tracking in Sheets**  
- Create/update a Google Sheet named "Meeting Preparation Tracker"
- Add columns: Date, Meeting Title, Time, Preparation Doc Link, Reminder Created (Yes/No), Status (Pending/Done)

7️⃣ **General Instructions**  
- Be concise, clear, and structured.
- Only include publicly available or non-sensitive content in the preparation.
- Respect privacy and do not share meeting content externally.
- Confirm completion by updating a doc named "Runner H Task Log".

**Goal:** Automate my weekly meeting preparation so I am fully prepared with minimal manual work, saving hours each week while ensuring effective participation in meetings.

---

### Testing Instructions:
✅ Ensure Runner H has access to:
- Google Calendar (read/write)
- Google Drive and Docs (create/edit)
- Google Sheets (create/edit)

✅ Use the current week for fetching events.

✅ Confirm by providing links to:
- The Weekly Overview Doc
- The Preparation Docs for each meeting
- The Preparation Tracker Sheet

✅ Ensure reminders appear in Google Calendar.

---

Here’s how I leveraged Runner H for end-to-end meeting preparation:

✅ Accessed Google Calendar: Pulled upcoming meetings for the week automatically.
✅ Generated Preparation Docs: Created structured Google Docs with meeting details, preparation points, questions to consider, and pre-reading suggestions using Surfer H.
✅ Pre-Reading Suggestions: Searched the web to add 3–5 relevant articles and summaries for each meeting.
✅ Scheduled Reminders: Added two automated Google Calendar reminders for each meeting (24 hours prior and 1 hour prior).
✅ Created a Weekly Overview Doc: Summarized all meetings for the week with preparation status and quick-access links to individual docs.
✅ Updated a Tracker Sheet: Added meeting details, preparation status, and links to a Google Sheet for progress tracking.
✅ Logged Completion: Documented the task in a Runner H Task Log.

To replicate:

Give Runner H access to your Google Calendar, Docs, Drive, and Sheets.
Use the system prompt I designed to run the automation weekly.
Watch your preparation workflow transform into a fully automated system.

Use Case & Impact

🛠️ Real-World Use Cases:

Knowledge workers, product managers, researchers, and students who need to prepare for meetings consistently.
Startup teams or solo founders managing busy calendars.
Anyone who wants to delegate repetitive preparation tasks to an agent to focus on high-leverage work.

Social Love

https://x.com/Kaustub88296852