NorthStar

Rendering UI
Workspace + Code
Tutorial Walkthrough
Timeline + Image Drag and Drop
Vibe Motion
Skills
Agents

Inspiration

Most science explainers lose people because they are either too abstract or too dense. I wanted NorthStar to make complex ideas feel clear in seconds by turning a rough prompt into a cinematic, structured visual story. The core motivation was educational impact: help students, educators, and creators move from curiosity to understanding faster. The barriers to entry dissolve! No more heavy rendering, difficult prompt engineering or the need to have a strong background in coding.

What it does

NorthStar is an agentic creator studio for Manim videos.

Converts one prompt into a scene-by-scene explainer pipeline: Plan -> Approve -> Code -> Render
Uses Gemini to generate a structured scene JSON with timings, goals, actions, and narration beats
Lets users edit the plan before code generation
Generates Manim Python from the approved plan and renders MP4 locally
Supports visual asset generation (BG/FG) and scene-level image mapping
Provides render diagnostics, auto-fix guidance, and retry flow when generation fails
Supports multi-session ideation, including parallel variants (Crazy mode)

UI + Backend Deployed: https://northstar-49nz.onrender.com (needs you own gemini api key) On the hosted app, Manim/ffmpeg are preinstalled in the deployment container; local install is only needed for self-hosting.

How we built it

We built NorthStar as a full-stack local-first application.

Backend (FastAPI + Python): job orchestration, Gemini integrations, Manim execution, ffmpeg post-processing, artifact management (work/jobs/<job_id>), health/status APIs
Frontend (HTML/CSS/JS): 3-panel IDE-style UI (Explorer, Canvas+Code, Chat+Agent), timeline editor, code editor, diagnostics panel, context memory/docs/skills
Gemini integration: planning, code generation, failure diagnosis, repair suggestions and code error fixes, source summarisation/youtube link indexing, nano banana pro image backgrounds, vibe motion text overlays.
Render system: deterministic job folders with plan/code/logs/video for reproducibility and debugging

Challenges we ran into

Keeping render reliability high when generated code can vary widely
Preventing UI overflow and layout breaks in 9:16 creator workflows
Designing timeline editing that stays simple but still precise
Avoiding noisy status spam while preserving transparency
Handling edge-case Manim errors fast enough for live demos

Accomplishments that we're proud of

Built a true end-to-end agentic video system: Prompt -> Plan -> Approve -> Code -> Render, instead of a single chat response.
Added a strong control layer with an explicit approval gate, so creators can edit structure before expensive rendering.
Implemented a practical self-healing loop: render error -> Gemini diagnosis -> fix guidance/diff -> retry, which keeps momentum during live demos.
Shipped scene-aware timeline operations with per-scene timing, notes, and visual mapping to keep storytelling coherent.
Added parallel variant generation (Crazy mode) to explore multiple creative directions from one idea quickly. This tests Gemini's reasoning capability to the limits!
Integrated multilingual narration/caption workflows to make outputs more globally usable.
Delivered a creator-first interface optimised for high-pressure demo use: transparent job artifacts, clear phase diagnostics, and fast pre-render iteration.

What we learned

Structured planning JSON significantly improves generation quality and consistency
Agentic loops (plan -> code -> repair) outperform one-shot generation for production use
UX clarity is as important as model intelligence in creative tools
Good defaults and visible artifacts reduce friction and build trust
“Fast to first result” is critical for creators and judges