I am an applied decision analyst building uncertainty-aware decision support systems using interpretable statistical and ML modeling, Bayesian inference, and uncertainty modeling in pharma, healthcare and industrial-support operations. My work emphasizes principled modeling, honest failure analysis, and designing analytical systems that make automated inference workflows reliable under uncertainity supporting direct human intervention and oversight.
Most data science systems optimize for predictive accuracy in static benchmark settings. My work instead focuses on decision-grade reliability in real operational environments characterized by sparse data and asymmetric failure costs.
Across projects, I design systems that:
- Explicitly model uncertainty rather than collapsing it into point predictions
- Encode operational constraints into decision rules
- Intentionally throttle automation when model confidence is insufficient
- Integrate human oversight as a first-class component of the system
- Treat model failure as a diagnostic signal about system structure, not a defect to be hidden
This perspective is informed by experience in regulated domains where false confidence, silent failure, and automation bias carry real-world consequences.
This project explores stochastic inventory forecasting under severe covariate scarcity using Poisson–Gamma conjugacy and a waste-constrained restocking policy.
It demonstrates both a principled Bayesian modeling approach and the structural limits of automated forecasting in nonstationary, human-driven consumption systems.
Key contributions:
- Exposure-normalized Poisson–Gamma modeling of consumption rates
- Closed-form posterior predictive forecasting via Negative Binomial distributions
- Monte Carlo decision policy optimizing reorder quantities under waste constraints
- Rolling-origin evaluation under regime shifts
- Formal falsification of automation viability under nonstationarity
Artifacts:
- Methods Note (PDF): https://thefifthpostulate.github.io/projects/stochastic-forecasting.html
- Analysis Notebook: https://thefifthpostulate.github.io/Stochastic-Consumption-Forecasting/InventoryProject.html
- Source Code: Available upon request
Key takeaway:
Uncertainty modeling revealed the true complexity of the consumption process. Assumptions about the stochastic process and decision rule were insufficient to consistently provide decision-grade forecasts, making expert oversight more reliable than fully automated inventory control.
Evidence Geometry is an experimental framework for interpretable risk modeling in classification problems. Instead of producing a single probability score, it decomposes model predictions into structured evidence signals that reveal how risk emerges in the data.
The framework transforms heterogeneous features into a unified log-likelihood ratio evidence space, allowing risk to be analyzed geometrically.
GitHub Repo
https://github.com/TheFifthPostulate/evidence-geometry/tree/main
Example Notebooks
Breast Cancer Wisconsin
https://thefifthpostulate.github.io/evidence-geometry/bcw_analysis.html
Cleveland Heart Disease
https://thefifthpostulate.github.io/evidence-geometry/heartdisease_analysis.html
Each feature contributes log-likelihood ratio evidence:
log p(x_i | positive class) − log p(x_i | negative class)
Stacking these contributions forms an evidence vector for each observation.
Working in this space provides several advantages:
- heterogeneous features become comparable
- evidence accumulates additively
- population structure becomes analyzable
- case-level risk can be decomposed
This connects the framework to classical likelihood ratio testing and Bayesian evidence accumulation.
For each case, the framework computes three complementary signals in the evidence space.
Difference in Mahalanobis distance to the learned class manifolds.
Measures which class distribution better explains the case.
Projection of evidence deviations from negative class manifold onto the mean class separation direction.
Captures net accumulation of evidence toward the positive class.
Energy along dominant covariance eigenmodes (principal component axes) of the positive class.
Detects activation of correlated pathological feature bundles.
Prototype v0.1
Current work focuses on:
- extending the framework to large clinical datasets (MIMIC-IV / eICU)
- exploring temporal risk signal evolution
- integrating triage policies based on evidence structure
Jithakrishna Prakash
📧 jprakashoff@gmail.com
🔗 LinkedIn: https://linkedin.com/in/jithakrishna-prakash
💻 GitHub: https://github.com/TheFifthPostulate