Cinematic Research

Inspiration

A filmmaker researching a creative idea rarely starts with a neat genre label. They might ask: “Which films make a city increasingly threatening over a single night—not merely use it as a backdrop? How do sound, lighting, and movement create that effect?”

A list of titles is only the beginning. The harder work is defining what actually fits, separating superficial similarities from meaningful references, and checking the evidence across languages. Cinematic Research makes that work inspectable.

What it does

Cinematic Research turns a nuanced film question into an editable research brief and an evidence-backed candidate ledger.

Define inclusion criteria, exclusions, and the evidence standard.
Inspect verified, borderline, and rejected candidates instead of receiving an undifferentiated recommendation list.
Open an evidence inspector to see support for individual claims, source links, and caveats.
Explore multilingual evidence with original passages alongside model translations.
Research cinematic lineage while distinguishing documented influence, critical comparison, and inferred similarity.
Revise the question and continue researching.

The intended users are filmmakers and screenwriters looking for creative references, film essayists investigating an argument, and fans exploring a precise cinematic interest.

How I built it

The application uses Next.js and TypeScript, with server-side Gemini calls on Google Cloud Vertex AI. Parallel is integrated at runtime through Gemini’s external grounding interface.

The research pipeline has four explicit stages:

Gemini compiles the question into a structured research brief.
A Parallel-grounded discovery call researches candidates and sources.
A separate Parallel-grounded verification call challenges the candidate set.
Gemini synthesizes the grounded material into the structured ledger and associates source references with evidence claims.

The server checks for grounding metadata rather than accepting an unsupported search result. Credentials stay server-side. The application has a standalone container configuration for Google Cloud Run.

There is no pre-populated film database: the research is assembled on demand from retrieved evidence.

Challenges I ran into

The central design challenge was making this more useful than another fluent film list. That meant exposing the research criteria, rejected candidates, and evidence gaps rather than hiding them behind polished prose.

Another challenge was keeping different kinds of claims separate. A critic comparing two films does not establish direct influence. A translated passage is not the original text. A verification label is a model assessment, not an objective guarantee. The interface keeps those distinctions visible.

Accomplishments

The working demo brings a complex question, a research brief, candidate evaluation, multilingual passages, and source inspection into one workspace. Parallel supplies the live research evidence, while Gemini plans, evaluates, and structures the result.

The public repository includes the application source, environment template, setup instructions, Dockerfile, and an MIT license. Lint, TypeScript checks, and a production build passed before publication.

What I learned

The useful output of film research is not just an answer; it is an argument someone can inspect and develop. Showing why a film almost fits can be as valuable as recommending one that does.

What's next

Improve source coverage, add stronger claim-to-source validation, and support more targeted research updates. Research can take several minutes, and coverage varies across languages and films. Model translations and classifications still need human scrutiny.

Try it

Start with a specific cinematic question, inspect the brief, select a candidate, and compare its evidence with the linked sources. For multilingual research, compare the original passages with the translations.

Source and local setup: https://github.com/varun-a-1010/cinematic-research