Probabilistic NLP · Responsible AI
DPULSE.AI — Probabilistic NLP for Organizational Narratives
Developed probabilistic NLP pipelines to identify emerging themes in employee narratives, with an emphasis on uncertainty calibration and responsible interpretation.
- Role
- Co-founder & Founding Engineer
- Institution
- Founded with two EPFL professors
- Date
- Sep 2025 – Jan 2026
Technologies
Context
Context
DPULSE.AI explored how probabilistic natural language processing can surface emerging themes in employee narratives. The emphasis throughout was on uncertainty calibration and responsible interpretation: treating model outputs as probabilistic evidence rather than verdicts, and being explicit about what the models can and cannot support.
Problem
Problem
Emerging themes in employee narratives are diffuse, evolving, and context-dependent, which makes them hard to capture with rigid, deterministic classifiers.
Outputs in this domain are sensitive, so calibrated uncertainty is essential to avoid over-confident conclusions.
Responsible interpretation requires keeping a human in the loop and resisting the temptation to over-claim what the model has detected.
Scope
System scope
Probabilistic topic modeling of employee narratives
Uncertainty calibration of model outputs
Responsible, human-in-the-loop interpretation
Product-oriented engineering scaffolding
Methods
Approach & methods
Used probabilistic topic modeling to represent emerging themes as distributions rather than hard labels.
Focused on uncertainty calibration so that confidence scores reflect genuine reliability.
Framed outputs around responsible interpretation, designed to support human judgment rather than replace it.
Built the engineering scaffolding needed to turn research ideas into usable tooling.
Contributions
Contributions
Co-founded the project and acted as founding engineer, building the probabilistic NLP pipelines and the engineering scaffolding around them.
Collaborated directly with two EPFL professors on the modeling and interpretation methodology.
Prioritized uncertainty quantification and responsible interpretation as first-class design constraints.
Results
Results
Role
Co-founder & Founding Engineer
Collaboration
2 EPFL professors
Focus
Uncertainty calibration
Process
Technical process
- 01Narrative corpus
- 02Probabilistic topic model
- 03Calibrated signals
- 04Human interpretation
Limitations
Limitations
Early-stage work: this case study deliberately makes no claims about traction, customers, accuracy, or deployment scale.
Theme detection in narratives is inherently subjective and benefits from careful human interpretation.
Labelled data in this domain is scarce, which constrains evaluation and motivates the focus on uncertainty.