AI-Powered PX: Sentiment & Analytics

The initiative was catalyzed by a critical bottleneck in strategic HR operations: the imbalance between data collection and data monetization. While the PX team successfully captured high volumes of organizational data (ranging from psychometric profiles to engagement surveys), the manual effort required to clean, normalize, and extract actionable insights delayed programmatic responses. The inspiration was to transition the HR function from a reactive operational unit to a predictive talent advisor by embedding intelligent automation directly into the workflow.

1. What We Learned Data integrity is the foundational prerequisite for any machine learning or automation layer. We learned that legacy spreadsheets and unstructured data models create immediate failures in automated pipelines. True operational efficiency is achieved not by building more complex models, but by enforcing strict data validation schemas at the point of ingestion and ensuring that downstream AI prompts receive clean, contextualized payloads.
2. How We Built It The architecture was built on a dual-layer approach: an Automation Layer for data pipeline management and an Intelligence Layer for data interpretation. We deployed automated workflows to handle repetitive scheduled triggers, parsing inputs dynamically from cloud spreadsheets. To evaluate the efficiency of our automated data sorting versus manual categorization, we utilized standard precision-recall evaluation metrics: Where \bm{T_p} represents true positives (correctly categorized talent/sentiment indicators), \bm{F_p} represents false positives, and \bm{F_n} represents false negatives. By optimizing the prompt engineering matrices within the LLM API calls, we pushed operational precision past a target threshold of 92%, significantly reducing the need for manual HR verification.
3. Challenges Faced Data Type Mutations: A significant technical hurdle involved automated tokens misinterpreting date formats generated by complex Excel formulas, converting them into unfilterable text strings. This was resolved by implementing strict data type formatting steps immediately post-ingestion. Contextual Anonymization: Maintaining strict data privacy while attempting to derive team-level sentiment analysis required building dynamic filtering matrices. The solution involved restricting the AI's data scope to aggregate cohort sizes (e.g., minimum sample size \bm{n \ge 5}) to prevent accidental identification of individual contributors.

Built With

• ai
• generatibe
• generative