Prompt Sensitivity MVP
What the Project Does
The Prompt Sensitivity MVP is a natural language processing (NLP) project that classifies short text inputs using machine learning techniques. The goal was to evaluate how well a model could identify patterns in text and make predictions based on language features.
The Challenge
While testing the model, I discovered that overall accuracy did not always reflect real-world reliability. Some short and ambiguous inputs were difficult for the model to classify correctly, highlighting the importance of trustworthy AI and human-in-the-loop decision making.
Example Failure Case
Input: “That’s crazy.”
Model Prediction: Strongly negative.
Why It Was Wrong: The phrase lacked context and could express excitement, surprise, or criticism depending on the situation.
AI Leaders Concepts Applied:
- Human-in-the-Loop Decision Making
- Trustworthy and Responsible AI
This example demonstrated that automated predictions should not always be accepted without considering uncertainty.
How I Solved It
Instead of focusing only on accuracy metrics, I analyzed failure cases and evaluated where human oversight would improve reliability.
I introduced the concept of confidence thresholds, where low-confidence predictions would be flagged for manual review rather than automatically accepted. This approach helps prevent incorrect classifications from being treated as definitive results and increases transparency in the decision-making process.
The project shifted from simply building a model to designing a more reliable and responsible system.
What It Taught Me
This project taught me that responsible AI development requires more than strong performance metrics. Building trustworthy systems means understanding model limitations, validating outputs, and incorporating human judgment when uncertainty exists.
The experience reinforced the importance of balancing automation with oversight and designing systems that prioritize reliability, transparency, and user trust.
Example confidence threshold logic
prediction = model.predict(text)
confidence = max(model.predict_proba(text)[0])
if confidence >= 0.80:
result = prediction
else:
result = “Flag for Human Review”
print(result)