Inspiration
Casting is one of the highest-stakes and least data-driven decisions in film.
The directors are unambiguous about how much it matters. Robert Altman: "When casting's done, 90 percent of my creative work is done." Martin Scorsese: "I always say that casting is 85 to 90 percent of the picture for me."
The people who actually do it are equally unambiguous that there's no method to it. Laura Rosenthal, casting director on Todd Haynes' films: "Sometimes it's just not a science, it's very instinctual." Jennifer Venditti (Euphoria), a nominee for the first-ever casting Oscar: "It's really hard to put words to it. It's like saying, 'Define beauty.'" And Rufus Sewell on chemistry: "Chemistry is nothing if it's there, but it's everything if it's not."
When instinct misses, it costs millions. Back to the Future shot five weeks of footage with Eric Stoltz before Zemeckis recast Marty McFly — the reshoot added $3–4M to a $15M budget, a ~25% overrun from a single casting call.
In 2026 the Academy awarded the first-ever Best Casting Oscar, its first new category since 2001. Hollywood has formally recognized casting as an art at the center of filmmaking — and it's still done entirely on gut.
If directing is 90% casting, why is casting still 0% data?
We aren't trying to replace instinct. We're trying to give instinct receipts: the documented collaborative history no human could synthesize per-decision, surfaced at the moment the decision is being made.
What it does
Upload a screenplay. Raphael reads it, decides who it needs, researches real people, and hands back a ranked set of complete casts you can edit live.
- Screenplay breakdown — drop a
.txtor.pdfscreenplay and Gemini parses it into a structured cast breakdown: characters, role presence (lead / supporting / ensemble), and the attributes each role demands. You review and edit the detected characters before anything else runs. - Candidate search — a casting-director agent fans out one concurrent candidate search per
character and returns a shortlist per role, each candidate carrying a role
fit_score. - Evidence enrichment — every unique candidate (deduped across roles) gets a dossier: filmography, documented collaborators and the specific projects they shared, real quotes about working together, awards, controversies, availability, union affiliation, languages and physical skills. Cached candidates come straight from ClickHouse; misses fall back to a live Parallel search.
- Risk assessment — a parallel risk agent scores each candidate Low / Medium / High against documented, sourced signals rather than vibes.
- Chemistry scoring — a deterministic engine builds a collaboration graph over every candidate and scores it. Pairwise edges combine NPMI and Adamic–Adar over shared production credits; pairs who've never worked together get a small, hard-capped 2-hop "shared collaborators" affinity so they rank against each other instead of flattening to zero. Per-actor terms fold in role fit and an asymmetric risk penalty. A density penalty deliberately discourages an all-star cast where every pair already has history.
- Cluster search — a 1-swap local search with random restarts, anchored on lead-role candidate combinations, returns the top complete casts rather than a single "best" answer.
- Live swaps — swap any actor and the score updates instantly.
/swap/previewscores every alternative in that character's shortlist first, so you see each candidate's chemistry delta and a per-co-star breakdown before committing. No agents, no API calls — a deterministic recompute over data the pipeline already produced. - Receipts on click — every score opens into the cited evidence behind it: the shared films, the real quote, the source.
- Per-user history — sign in with Google and every screenplay and report is saved to your own project history, resumable later.
How we built it
An agentic pipeline on Google Cloud (Gemini via Vertex AI), a factual backbone from Parallel, and an OLAP layer in ClickHouse — each doing exactly the one thing it's best at.
screenplay ──► ScreenplayBreakdownAgent ──► CastingDirectorAgent ──┬─► EnrichmentAgent (ClickHouse ─► Parallel)
(Gemini) (fan-out search) └─► RiskManagementAgent
│
ChemistryEngine ◄────┘
(graph + local search)
│
RecommendationEngine ──► RecommendationReport
- Gemini (Vertex AI) is the agentic brain: parsing screenplays into structured objects with constrained output, and reasoning over Parallel's facts into chemistry inputs and risk assessments.
- Parallel is the factual backbone. We ask it only for what it's best at — discrete, citation-backed facts (filmography, documented collaborations, real quotes, availability) through its Basis sourcing framework. We deliberately never ask it for subjective judgments. Raphael computes chemistry itself, on top of Parallel's verified evidence. Deep research is reserved for the offline ETL; the online path uses shallow search only.
- ClickHouse is the core loop, not a side dashboard. Dossiers are stored as a wide,
array-typed table and the roster search-and-rank across every researched candidate runs as an OLAP
query at request time. Reached over MCP (
mcp-clickhouse) through a single long-lived session opened at app startup, so requests don't each pay for a subprocess and schema setup. Live on-screen swaps stay local, respecting ClickHouse's own anti-pattern guidance on point lookups. - Agents are LangGraph
StateGraphs. Every fan-out (per character, per candidate) usesSendinto an async branch node and a reduce node that builds the report — so a five-character screenplay researches five shortlists concurrently, not serially. - Backend: FastAPI + Pydantic,
uv-managed, deployed to Cloud Run from GitHub Actions./recommend/streamstreams stage-by-stage progress so the UI isn't a spinner. - Frontend: React + TypeScript + Vite. Ingest, Roster (formation tree, ensemble chemistry graph, cluster tabs, risk register) and About screens. Firebase Auth for Google sign-in, deployed to Firebase Hosting.
- History: Cloud Firestore under
users/{uid}/projects, written only from the backend via the Admin SDK —firestore.rulesdenies all direct client access. - Offline ETL: a TMDB-seeded discovery CLI walks real films' top-billed casts, runs Parallel deep research per person, and populates ClickHouse ahead of time, so the demo path is fast and the candidate universe is real.
Challenges we ran into
- Defining "chemistry" without hand-waving. The naive version — count shared credits — rewards prolific actors and says nothing about pairs who've never met, which is the overwhelmingly common case. We landed on NPMI plus Adamic–Adar over the collaboration graph, with a small, capped 2-hop shared-collaborator backoff so unproven pairs still differentiate. Every weight is tunable via env var because we retuned constantly against real screenplays.
- Stopping the model from grading its own homework. Early versions asked an LLM to just rate chemistry, and it produced confident, unfalsifiable numbers. Splitting the responsibilities — Parallel gathers only citable facts, Raphael computes the score deterministically — is what makes "click for receipts" possible at all.
- The all-star trap. The first working scorer converged on casts where everyone had already worked with everyone, which is both boring and unrealistic. We added an explicit density penalty toward a target of ~0.35 to push the search back toward mixed casts.
- Latency. A naive pass was one deep-research call per candidate per role — minutes per
screenplay. Fixing it took four things: deduping candidates across roles, LangGraph
Sendfan-out so branches run concurrently, ClickHouse-first lookups with deep research pushed offline into the ETL, and streaming progress to the UI. - Keeping one MCP session alive. ClickHouse is reached through an
mcp-clickhousestdio subprocess, which can't be opened in a synchronous constructor and can't be reopened per request without paying subprocess and schema-setup cost every time. It's opened once in FastAPI's lifespan handler and shared by every request — which also drove auv-at-runtime requirement into the Cloud Run image. - Real screenplays are messy. PDFs of actual scripts are inconsistently formatted; getting a reliable character breakdown out of them needed structured output plus a human review step in the UI rather than blind trust.
Accomplishments that we're proud of
- Swaps are instant. The expensive pipeline runs once; every subsequent swap and preview is a deterministic local recompute over the roster already in hand. Editing a cast feels like a game, not like waiting on an API.
- Every number opens into its evidence. No score in the product is unexplainable — you can click through to the films, the quote, and the source.
- We ship complete casts, not a leaderboard. The cluster search returns several coherent ensembles, which is how the decision is actually made.
- A real, end-to-end product. Auth, per-user history, streaming progress, formation tree, ensemble chemistry graph, risk register, report export, CI/CD to Cloud Run and Firebase Hosting — not a notebook demo.
- Honest division of labor across three services. Parallel gathers verifiable evidence, ClickHouse ranks at scale, Gemini reasons over it. Nothing is a name-drop.
What we learned
- Ask each service for what it's uniquely good at, and nothing else. The quality jump came from narrowing what we asked of the LLM, not from asking it for more.
- Deterministic scoring is a feature, not a compromise. Because chemistry is computed rather than generated, it's reproducible, tunable, explainable, and fast enough to recompute on every keystroke.
- Graph metrics from information retrieval transfer surprisingly well. NPMI and Adamic–Adar, borrowed from co-occurrence and link prediction, capture "these two have real history" far better than raw counts.
- Caching is architecture. Deciding what is offline (deep research) versus online (search and rank) determined the entire shape of the system.
- Users need a checkpoint. Letting people correct the screenplay breakdown before the expensive pipeline runs cost one screen and bought all of the trust.
What's next for Raphael
- Beyond actors — extend chemistry to directors, writers, cinematographers, and the rest of the crew. The graph and scoring are role-agnostic already; it's a data-coverage problem.
- Richer chemistry signals — genre fit, skills, and critical reception of past collaborations, not just whether two people have shared a set.
- Real-world constraints — budget, scheduling, and availability, so a great cast on paper is one that can actually be assembled. Projected hours per cast member against a production timeline, and calendar conflict detection across the roster.
- Opening the door for newcomers. This is the direction we care about most. A tool that scores people by documented history structurally favors people who already have one — which is exactly the loop that keeps first-timers out of the room. The density penalty is a first step: it already pushes the search away from casts where everyone has worked with everyone. Next is going further — an explicit discovery mode that surfaces high-fit, low-credit candidates alongside the safe picks, and a chemistry model that predicts affinity from shared collaborators, training, and background rather than requiring a shared credit to exist at all. The goal is to make taking a chance on someone new a defensible decision with evidence behind it, instead of a leap of faith that nobody wants to be responsible for.
- Casting directors keep the final call. Raphael arms them; it doesn't overrule them. The industry just put casting on its biggest stage — we want to be the first tool that treats it with the rigor that recognition demands.
Why Raphael?
Named after the Great Sage — later known as Raphael — from That Time I Got Reincarnated as a Slime, this project draws from what makes that character compelling: not omniscience, but tireless analysis, precise reasoning, and the ability to act autonomously on behalf of its user.
Log in or sign up for Devpost to join the conversation.