AI summary by TwinMind
Human Stories in AI: Achal Dixit
StatQuest with Josh Starmer
33m 12s
Introduction and Background
- Josh Starmer hosts Human Stories in AI podcast with guest Achel Dixit
- Achel: Data scientist at Delivery (largest fully integrated logistics services in India)
- Previous roles: Business Technology Analyst at ZS, Research Assistant at Imperial College London
- Childhood interest in computer science, ML, and AI led to computer science engineering degree
COVID-19 Hackathon and First Research Publication
- First major breakthrough during COVID lockdowns in second year of engineering
- Participated in MIT COVID-19 hackathon with 800+ participants
- Team won hackathon working with University of Michigan researchers
- Problem: Predicting ICU times for COVID patients
- Pivoted to predicting COVID severity based on demographics, symptoms, comorbidities
- Used logistic regression to classify severe vs. non-severe cases
- Published research in American Journal of Emergency Medicine
- Goal: Help patients understand precautions needed based on risk factors
Heart Failure Research Project
- Third year machine learning course project
- Identified gray area in heart failure diagnosis using ejection fraction (40-50% range)
- Found problem through literature review of clinical papers
- Used public dataset from PhysioNet
- Applied exploratory data analysis and visualization techniques
- Used PCA, UMAP for dimensionalization
- Plotted creatinine and other features vs. ejection fraction
- Hypothesis: Linear scale metric flaw causing misclassification
- Converted to binary classification using semi-supervised learning and active learning
- Published "Demystifying Heart Failure with Mid-Range Ejection Fraction" at 48th Computing in Cardiology Conference
Learning Philosophy and Approach
- Describes approach as "exploration vs. exploitation dilemma" from reinforcement learning
- Balance between learning new things (exploration) and applying known tools (exploitation)
- Evidence-based stepping: move forward only with sufficient evidence
- Focus on visualization first to understand data patterns
- Problem-solving approach: find problem → quantify methodology → run experiments
Microsoft Imagine Cup and Product Development
- Built EHR (Electronic Health Records) application for Indian healthcare
- AI algorithms run in background providing recommendations to clinicians
- Reached top 4 finalists out of 1000+ participants at Microsoft Imagine Cup
- Product addressed lack of EHR adoption in India
Industry Philosophy and Methodology Selection
- Key insight: "People get obsessed with methodology and outcome rather than impact"
- Factors beyond accuracy: cost, training time, maintenance, team learning curves
- Example: TabNet vs. XGBoost comparison
- TabNet: 5% better accuracy, requires GPU, expensive cloud compute
- XGBoost: Faster training, lower cost, easier maintenance
- Good data scientist translates business problems to mathematical problems
- Uses optimal tools without incurring excessive costs
Current Work at Delivery
- Logistics optimization using Mixed Integer Linear Programming (MILP)
- Problems solved: facility location, rider assignment, vehicle routing, load assignment
- MILP preferred for explainability (white box vs. black box methods)
- Uses statistical machine learning on tabular data
Ice Cream Business Analogy for Logistics Problems
- Benchmarking problem: determining optimal performance for each location
- Variables: city differences, product mix (100 flavors/brands), salespeople, locations
- Two main challenges:
- How to benchmark performance
- Identifying causality for underperformance
- Solution approach:
- Collect historical data on people, places, sales figures
- Use Uber H3 hexagon mapping for locality features
- Create features: locality preferences, brand performance, customer response
- Train tree-based model on patterns
- Apply SHAP (Shapley) analysis for feature contribution
- Key assumption: training data represents optimal performance
- Compare current performance against model predictions
Data Science Reality
- 99% of data science involves getting good data
- Modeling is only 10% of the work
- Major challenges: finding proxies for missing data, making reasonable assumptions
- Extensive exploratory data analysis and hypothesis testing
- Example: Testing affluence correlation with ice cream sales in rich localities
- Validate hypotheses against historical data before including in models
