Human Feedback in AI: How Unit8 Approaches Validation for Reliable AI Systems

The integration of human feedback in AI systems is critical for ensuring reliability and performance. At Unit8, we have observed that effective validation processes often hinge on domain expertise rather than the use of dedicated validation teams. Drawing from our extensive experience across industries, we’ve developed insights and best practices for integrating human feedback into AI workflows.

AI Validation: Insights from Unit8’s Experience

Unlike large-scale AI initiatives like OpenAI’s development of ChatGPT, which relied on contractors for Reinforcement Learning from Human Feedback (RLHF), our approach leverages existing teams and subject matter experts (SMEs) to validate AI outputs. This ensures that the nuanced understanding required for specialized domains, such as insurance or regulatory compliance, is always present.
While external validation teams may be effective for generic tasks, their limitations become apparent in industries like healthcare, insurance, and manufacturing, where domain-specific knowledge is essential.

Key Project Highlights

1. Human-in-the-Loop (HITL) for Insurance AI Models

• Challenge: Processing 20 million claims for a reinsurance company’s property and casualty division.
• Approach: We designed a sampling-based validation process integrated into Palantir Foundry to ensure efficient workflow management.
• Key Insight: SMEs, while invaluable for validation, often face resource constraints. This underscores the importance of allocating qualified personnel and establishing standardized processes.

2. UI-Based Validation for Regulatory Compliance

• Challenge: Extracting regulatory and recipe information from hundreds of documents for a Swiss flavor and fragrance manufacturer.
• Approach: A two-phase pipeline was developed:

1. 1. Data extraction into a database with a regulatory-compliant model.
   2. A side-by-side UI for human validation, ensuring accuracy and compliance.

• Key Insight: Human validated examples provided as part of few-shot prompts improve extraction much more than other prompt adjustments. Selecting the right examples for a given document to extract is crucial for accurate results.

3. Multi-Layer Validation for Medical Device Documentation

• Challenge: Improving manuals by extracting data from two decades of support tickets with current throughput reaching approximately 10,000 annually.
• Approach: A two-layer validation process:
  • Initial review by non-experts (300 staff members).
  • Final validation by senior experts (10 specialists).
• Key Insight: This structured approach balanced expertise and workload, proving effective for large-scale validation.

4. Systematic Feedback Integration in E-Commerce

• Challenge: Refining an AI pipeline for faster, more accurate results.
• Approach: Using the Langfuse framework, we systematically collected and integrated human preferences to fine-tune a smaller, efficient model.
• Key Insight: Human feedback not only validates outputs but can also enhance system efficiency by training models to align with user expectations.

Best Practices for AI Validation

From these experiences, we’ve identified several best practices:

• Domain Expertise Matters: Validation requires subject-specific knowledge to ensure high-quality results.
• Integrated Workflows: Tools like Palantir Foundry and custom UIs streamline validation within existing processes.
• Layered Validation: Combining non-expert reviews with expert oversight balances accuracy and resource management.
• Sampling Strategies: For large datasets, validating a representative subset can yield reliable insights while saving time.
• Feedback for Improvement: Frameworks like Langfuse turn validation into a tool for pipeline optimization.
• Separate Datasets: Maintaining clear distinctions between raw and validated datasets ensures traceability and quality control.

Reimagining Validation with Unit8

At Unit8, we believe that successful AI validation comes from involving existing teams and experts in the feedback process rather than creating separate validation teams. This approach reduces costs, utilizes in-house expertise, and ensures that AI systems align closely with business objectives.

As industries increasingly adopt AI, these insights and practices can serve as a foundation for building robust, reliable validation processes that drive innovation while maintaining quality.

Are you ready to optimize your AI workflows with a tailored approach to validation? Let Unit8 guide you on the path to smarter, more reliable AI systems.