More natural text with critical facts preserved
The prototype rewrites text for more natural tone while preserving protected names, numbers and facts.
Small language model case study
Humanizer SLM
Constraint-first text rewriting that improves naturalness while preserving names, numbers and critical facts.
Modern content teams, SaaS platforms, and digital agencies are generating an increasing volume of written output with the aid of large language models. While AI accelerates content production, it introduces a recognizable pattern in prose stilted phrasing, templated structure, and a mechanical rhythm that sophisticated readers, search algorithms, and platform detectors can identify instantly.
Project classification: This prototype explores controlled text rewriting that improves tone and readability while preserving protected facts and entities.
Less manual proofreading and correction
Automated comparison reduces the amount of manual proofreading needed to identify unintended changes.
Configurable tone and domain behaviour
Tone and domain controls support different communication styles and use cases.
01 / Business context
Who needed the solution
Modern content teams, SaaS platforms, and digital agencies are generating an increasing volume of written output with the aid of large language models. While AI accelerates content production, it introduces a recognizable pattern in prose stilted phrasing, templated structure, and a mechanical rhythm that sophisticated readers, search algorithms, and platform detectors can identify instantly. The client required a reliable and repeatable solution to transform AI-generated or robotic text into natural, human-quality prose without compromising factual accuracy or domain integrity.
02 / Challenge
What needed to change
Rewriting must be sufficiently adaptive to sound human, yet strictly constrained to preserve factual accuracy. Balancing these opposing requirements forms the core engineering challenge addressed by Humanizer SLM.
AI-generated text often sounded repetitive or roboticAI-generated text is increasingly detectable by humans, search engines, and content screening tools, reducing trust and engagement.
Generic rewriting tools changed names, numbers or URLsGeneric rewriting tools introduce errors by altering names, numbers, URLs, and key facts, making outputs unreliable for professional use.
Prompting alone did not reliably preserve critical factsRewriting must be sufficiently adaptive to sound human, yet strictly constrained to preserve factual accuracy.
Teams needed configurable tone and domain behaviourContent teams lacked tone control across different communication styles such as professional, friendly, persuasive, and concise.
Editors needed transparency and rollbackNo audit trail existed to track changes, review differences, or revert to previous versions.
03 / Workflow transformation
Before and after
Previous workflow
- AI-generated text often sounded repetitive or robotic
- Generic rewriting tools changed names, numbers or URLs
- Prompting alone did not reliably preserve critical facts
- Teams needed configurable tone and domain behaviour
- Editors needed transparency and rollback
Structured workflow
- Fine-tune a focused rewriting model
- Lock critical entities
- Configure tone and domain
- Validate and compare versions
- Review structured outputs
04 / Solution
How we approached it
The solution is built on a fine-tuned Google Flan-T5 model a lightweight, instruction-following sequence-to-sequence transformer selected for its efficiency and precision in constrained text transformation tasks. Model Selection: Google Flan-T5 Flan-T5 was selected for the following reasons: 1.
Fine-tune a focused rewriting model
Adapt Flan-T5 to a constrained text-transformation task.
Lock critical entities
Protect names, numbers, URLs and domain-sensitive phrases during generation.
Configure tone and domain
Control style through reusable templates rather than separate models.
Validate and compare versions
Check factual integrity and expose differences, history and rollback options.
05 / Example workflow
From input to reviewable output
1Fine-tune a focused rewriting model
2Lock critical entities
3Configure tone and domain
4Validate and compare versions
5Compare versions and protected facts
6User approval
06 / Delivery scope
What the delivery covered
Rewriting workflow
- Input text and protected-entity capture
- Tone and domain configuration
- Versioned rewrite generation
- Diff and factual-preservation review
Model and guardrails
- Flan-T5 fine-tuning with LoRA
- Fact and entity locking
- Constraint-based output validation
- Quality and mutation testing
Application engineering
- React interface
- FastAPI service layer
- Pydantic schemas
- Python processing and comparison pipeline
Scope boundary
The prototype is positioned around clarity, tone and factual preservation, not evasion of integrity controls. No zero-mutation claim is made without a documented test method.
07 / Architecture and controls
How the system is organised
Flan-T5
React
Python
FastAPI
Pydantic
PyTorch
Pandas
LoRA
Protected-entity locking
Names, numbers and selected domain facts are locked before generation and checked afterward.
Diff and version review
Users can compare versions and inspect changes before accepting the rewrite.
Acceptable-use boundary
The product should be positioned around clarity and tone, not bypassing integrity or detection controls.
Factual-preservation testing
Any zero-mutation claim requires documented test coverage and methodology.
08 / Business value
Value observed during validation
1Pilot-observed
More natural text with critical facts preserved
The prototype rewrites text for more natural tone while preserving protected names, numbers and facts.
2Pilot-observed
Less manual proofreading and correction
Automated comparison reduces the amount of manual proofreading needed to identify unintended changes.
3Pilot-observed
Configurable tone and domain behaviour
Tone and domain controls support different communication styles and use cases.
4Pilot-observed
Transparent changes through diff and version history
Constraint checks reduce factual mutations introduced by unconstrained rewriting.
5Pilot-observed
Lower-cost deployment through a focused small language model
Version history and diff review make editorial changes easier to inspect before approval.
09 / Technology
Every component, with its role in the delivery.
Technology choices from the supplied project brief, mapped to the workflow each component supports.
Flan-T5
Fine-tuned text-rewriting model
React
Reviewer or product user interface
Python
AI, data-processing and backend logic
FastAPI
Backend APIs and workflow orchestration
Pydantic
Structured schema and output validation
PyTorch
Model training and inference
Pandas
Tabular data processing and analysis
LoRA
Parameter-efficient model fine-tuning
NumPy
Numerical processing
Related work
Explore related case studies.
