Editor de Libros con LLM - Revisor Local de Manuscritos

Descripción

I created the local manuscript reviewer to solve a practical limitation of general-purpose language-model interfaces: a book is too large and structurally complex to treat as one prompt, and an unpublished manuscript may be too sensitive to send through an uncontrolled workflow. The project turns editorial review into a repeatable local pipeline. It accepts a long document, divides it into meaningful working units, builds searchable context, and performs focused review passes while keeping the author in control. The design reflects my experience with secure AI, document processing, data residency, and workflow automation rather than treating model access as the entire product.

The first architectural problem was context management. Splitting a manuscript at arbitrary character limits can separate a claim from its evidence, a scene from its setup, or a pronoun from the subject it references. I designed the ingestion process around document structure and context-aware chunking, preserving useful boundaries while keeping chunks small enough for local inference and retrieval. Metadata records where each segment belongs in the book, allowing later stages to reconnect a passage with its chapter and surrounding material. This is the same contract-first thinking I use in distributed systems: preserve identity, ordering, and provenance as information moves between components.

Embeddings and local search provide the reviewer with selective memory. Instead of repeatedly loading the entire manuscript, the system can retrieve passages related to a character, topic, term, or editorial concern. My work with vector search, RAG, TorchSharp, and machine-learning systems helped me separate retrieval quality from generation quality. A fluent answer is not necessarily grounded, so the workflow keeps the retrieved source material visible and treats it as evidence for a suggestion. SQLite stores document metadata and processing state, giving the pipeline a simple, inspectable record of what was indexed, reviewed, and produced.

I organized the editorial process as multiple deliberate passes rather than one vague request to improve the book. A pass can focus on continuity, repetition, clarity, tone, terminology, pacing, or structural consistency. Each pass has a defined objective, input scope, and output format, which makes results easier to compare and rerun. This approach draws on software architecture and agentic AI principles: decompose a broad goal into bounded tasks, give each task only the tools and context it needs, and retain checkpoints between steps. The model proposes; the workflow coordinates; the author remains the final decision-maker.

Privacy influenced every layer. Local execution reduces unnecessary disclosure, but local does not automatically mean secure. The application still needs controlled file access, safe temporary storage, clear retention behavior, and protection against one manuscript being mixed with another. I applied knowledge from security engineering, identity and access management, cryptography, and compliance-oriented architecture to define trust boundaries around the files, index, model process, and exported recommendations. The project does not claim to replace a formal security program; it demonstrates how privacy requirements can shape an AI system before features and convenience make those decisions difficult to change.

Operational reliability matters even for a single-user tool because manuscript processing can be expensive and time-consuming. The pipeline records progress, isolates stages, and supports repeatable execution so a failure does not require restarting every completed step. Docker provides a consistent runtime, while local persistence makes behavior easier to inspect and recover. I used the same resilience ideas found in larger cloud platforms: idempotent work where practical, explicit state transitions, bounded retries, useful diagnostics, and outputs tied to their inputs. These controls turn an AI experiment into a tool that can be used repeatedly with understandable behavior.

My study of Model Context Protocol and autonomous AI architecture also provides a path for evolving the project without collapsing everything into a single agent. File readers, retrieval services, style guides, and export tools can be exposed through narrowly defined interfaces, while an orchestrator selects them under explicit policy. That extension would require permission boundaries, audit records, and evaluation before being trusted with autonomous changes. The current implementation therefore emphasizes assisted review rather than unattended rewriting. It uses the advantages of language models while respecting uncertainty, authorship, and the need to trace a recommendation back to manuscript evidence.

This project brings together my experience in C# and .NET, local data systems, deep learning, vector retrieval, secure architecture, and technical writing. Its most important outcome is not a claim that an LLM can edit a book alone. It is a practical demonstration that AI quality depends on the system around the model: how context is prepared, how evidence is retrieved, how tasks are bounded, how state is recorded, and how a human evaluates the result. By building those controls into the workflow, I was able to apply modern AI capabilities to a real creative process without sacrificing privacy, repeatability, or editorial ownership.