← Home
ScriptSense logo

Production AI + Entertainment Intelligence

ScriptSense

An agentic AI platform that transformed entertainment IP into structured, searchable intelligence.

I led architecture and engineering as ScriptSense evolved from an early script-analysis product into a multi-workflow production AI platform.

The ScriptSense workspace — structured analysis alongside editable narrative output.

The ScriptSense workspace — structured analysis alongside editable narrative output.

I led the architecture and engineering of ScriptSense, a production AI platform built to help entertainment teams analyze, compare, search, and work with scripts, books, and other intellectual property.
The platform combined retrieval, model orchestration, structured extraction, specialized AI workflows, and human review to support script coverage, draft comparison, production breakdowns, content discovery, and conversational access to large libraries.
What began as a focused script-analysis product evolved into a broader platform designed for repeatable entertainment-intelligence workflows.

Project

Company

Jumpcut Media

Industry

Entertainment technology

Product stage

Early product & idea → production platform

Responsibility

Role

Head of Engineering & Lead Architect

Engagement

Full-time engineering leadership

Scope

Architecture, engineering leadership, AI workflows, retrieval, infrastructure, model strategy, and production delivery

Team

Led engineering and technical direction across a cross-functional group spanning product, software, AI/ML, and entertainment-domain expertise

System & outcome

Technology

Application — React, Next.js, TypeScript, NestJS (Node.js)

AI — Multi-step LLM pipelines and agent orchestration, critique loops, provider failover, RAG & vector retrieval, structured extraction, model routing

Platform — AWS S3, Postgres, MongoDB & Pinecone, durable jobs (Inngest), Valkey, tenant-level permissions

Models — Anthropic (Claude), OpenAI, Voyage embeddings

Outcome

Jumpcut Media acquired by Cinelytic (2025)

The problem

Entertainment organizations hold enormous volumes of valuable intellectual property, but much of the intelligence inside those assets remains trapped in unstructured text. Workflows such as script coverage, draft comparison, production breakdowns, catalog research, and content discovery are slow, specialized, and difficult to scale.

The opportunity was not to replace creative judgment. It was to give development and production teams a faster, more consistent way to surface evidence, structure information, compare material, and work across large content libraries without abandoning the processes they already trusted.

My role

I led engineering and served as the platform’s primary architect from early product development through production scale. I owned technical direction across the application, AI pipelines, retrieval systems, workflow orchestration, infrastructure, model strategy, and engineering delivery.

I was Head of Engineering and lead architect, and I stayed hands-on throughout. I owned the architecture and wrote much of the implementation myself: agent orchestration, APIs, data pipelines, and the Next.js and TypeScript frontend, alongside tenant-level permissions for enterprise clients.

I also led the practical decisions required to turn an ambitious AI concept into a production platform: sequencing the roadmap, controlling inference cost, managing latency, evaluating model behavior, building recoverable long-running jobs, and ensuring the architecture could expand beyond its original use case.

System architecture

Content ingestion and normalization

Scripts, books, and related content entered through ingestion workflows, where documents were parsed and normalized into a representation that supported both deterministic processing and model-assisted analysis. Long-form documents required contextual chunking that preserved structure.

Intelligence and retrieval

Structured data supported repeatable downstream workflows, while vector retrieval — using metadata and document structure — surfaced relevant passages, so entire documents did not need to be loaded into every model call.

Workflow orchestration

Distinct workflows handled different tasks. Long-running jobs required state, retries, progress tracking, and recoverability, and the system coordinated retrieval, model calls, validation, and persistence.

Model abstraction

The system was not tied to one model provider; models could be selected by quality, context, latency, and cost, and stronger models could be adopted without rebuilding the product.

Application and tenancy

Outputs were persisted and revisitable. Users needed projects, permissions, and shared workflows, and the system was designed for organization-level use.


A simplified view of how long‑form entertainment content moved from ingestion through retrieval, workflow orchestration, validation, persistence, and human review.

The AI pipeline

ScriptSense did not treat every analysis task as the same prompt applied to a different document. Each workflow had different output requirements, context needs, evaluation criteria, and tolerance for ambiguity.

Script coverage required a coherent editorial assessment. Draft comparison required alignment across versions and evidence tied to specific changes. Production breakdowns required structured extraction that could support downstream planning. Content discovery drew on vector retrieval across the analyzed library, while conversational access grounded its answers in the structured analysis the platform had already produced.

The orchestration layer coordinated specialized model calls, deterministic processing, retrieval, validation, and persistence, so each workflow could be optimized independently while still sharing a common platform foundation. The important achievement was not merely running many model calls — it was designing the system around the distinct requirements of each workflow.

Script coverage workspace combining structured analysis with editable narrative output.

Production considerations

Getting the system into production required more than making model calls succeed. Reliability, evaluation, evidence, cost, latency, recoverability, and human review all had to become first‑class architecture concerns.

Reliability


Long‑running jobs needed persistent intermediate state, retry and recovery behavior, and safeguards against silent downstream corruption.

Evaluation


Completion did not equal quality. Outputs had to be assessed for completeness, evidence, consistency, and usefulness — and each workflow needed its own evaluation criteria.

Cost and latency


Long documents and multi‑stage workflows were expensive. Models were selected by task, lower‑cost models were used where quality held up, and parallelism was controlled.

Human review


The system accelerated experts rather than replacing them. Outputs stayed inspectable and editable, and evidence and structured data supported validation.

Model independence


Provider abstraction reduced vendor dependency; routing could weigh quality, cost, latency, and context limits, and the platform could adopt stronger models over time.

Key decisions

Matching each task to the right model

Constraint: Long scripts and multi-stage workflows made inference the largest variable cost, and not every call needed the strongest model. Editorial analysis needed depth; keeping a chat’s running memory needed speed and a low price.

Options considered: One top model for everything, which is simple and expensive; the cheapest model everywhere, which weakens the analysis; or choosing the model per task behind a shared provider layer.

What I chose: Per-task selection. Analysis and structured extraction run on Claude Sonnet, lighter work like conversation memory runs on Claude Haiku, and retrieval uses Voyage embeddings, all behind an abstraction that isn’t tied to one provider.

What happened: Lower-cost models took the work where quality held up, and the expensive calls went where they changed the result. When stronger models shipped, each workflow could adopt them on its own schedule without a rebuild.

Critique loops and provider failover

Constraint: A first-pass model answer on a long script is often plausible but uneven, and coverage readers notice a weak synopsis or a missed character at once. Model providers also have outages and rate limits, and a stalled provider can’t mean a stalled report.

Options considered: Accepting the first output and relying on human review to catch problems, adding a second model to review the first, or tightening prompts alone. For availability: retry the same provider, or route to another one.

What I chose: Critique loops on the stages where quality varied most: one pass drafts, another checks the draft against the script and sends it back with specific objections before it is saved. Failover sits in the same provider layer as model selection, so a failing call moves to an equivalent model without the workflow knowing.

What happened: Weak drafts were caught before a reader saw them, and provider trouble stopped being a user-facing event. Because both behaviors lived in the shared layer, every workflow got them without its own code.

Workflow‑specific systems

Script coverage, draft comparison, production breakdowns, and content discovery required different context, validation, and output structures. A single universal agent would have hidden those differences rather than handling them well.

Structured intelligence alongside narrative output

ScriptSense preserved queryable data in addition to generated prose, so analysis could support later workflows, search, comparison, and product features.

Durable, step-based jobs for long-form analysis

Constraint: A full coverage report on a feature-length script is a long chain of model calls: synopsis, beat sheet, characters, content rating, casting recommendations, comparable titles, then the finished report. A failure near the end couldn’t mean starting over. That wastes the user’s time and pays for every completed call twice.

Options considered: Running the analysis inside the request, a plain job queue that retries the whole job, or a durable workflow engine that checkpoints each step.

What I chose: Durable functions on Inngest. Each stage runs as its own step, and its output is saved before the next one starts. Heavier side work, like per-character dialogue stats, goes out as a separate event so it doesn’t hold up the report.

What happened: A failed stage retries on its own and the job resumes from the last completed step, so a timeout late in the run reruns one step instead of the whole report. Later work, including embeddings and search backfills, reused the same pattern rather than getting its own job system.

A shared design system for nontechnical users

The people using ScriptSense were development executives and readers, not engineers, so the interface had to explain itself. I built a design system of more than 40 reusable components in Next.js and TypeScript. New workflows assembled screens from it instead of starting over, which cut UI development time by 35%.

Human review as part of the product

The system accelerated expert judgment rather than attempting to replace it. Outputs needed to be inspectable, editable, and tied back to supporting evidence.

Outcome

Coverage turnaround cut from a week or longer to under five minutes.

From focused script analysis to a reusable production AI platform — built before Jumpcut Media’s 2025 acquisition.
Over 100,000 hours saved across four years in production, which comes to an estimated $4 million in labor at a $40/hour floor. That rate is deliberately low: it sits at the bottom of what a fully loaded analyst or coordinator costs, so the real figure is likely higher.
ScriptSense began as a focused script-analysis product and evolved into a broader entertainment-intelligence platform supporting analysis, comparison, production planning, discovery, and conversational access across large content libraries.
Its shared ingestion, retrieval, orchestration, and persistence layers allowed new entertainment workflows to be introduced without rebuilding the technical foundation for each use case.
The platform became a central technical asset of Jumpcut Media before the company’s acquisition by Cinelytic in 2025.
The shared design system, more than 40 reusable components, cut UI development time by 35%, so each new workflow shipped with a consistent interface instead of a one-off build.

What this work taught me

ScriptSense reinforced a distinction that now shapes how I approach production AI: successfully executing a workflow is not the same as producing a trustworthy decision.

Logs could show that models ran, retrieval completed, and outputs were generated. They could not automatically tell us whether the system used the right context, resolved ambiguity appropriately, or produced a result a domain expert should trust.

That gap made evaluation, evidence, model routing, human review, and behavioral visibility first-class architecture concerns rather than features to add after launch.

Testimonial

Built and maintained our platform. Well organized, delivers the development roadmap on time. They're a highly skilled developer with a strong work ethic.

Tobias Queisser · CEO, Cinelytic

Skills

AI systems architecture · AI workflow orchestration · Retrieval‑augmented generation · Long‑context document processing · Model routing · Production AI evaluation · Technical leadership · Multi‑tenant platform architecture · Background job systems · Human‑in‑the‑loop workflows · JavaScript · Mobile · NEXT.js · NodeJs

Featured in

Building a production AI platform with complex workflows, long-form content, or high-stakes outputs? This is the kind of system I architect and lead.

Book a call →