Himanshu
Yadav

Problem

Fraud patterns evolve continuously, making static detection systems unreliable in production.

Approach

Streaming-first anomaly detection with monitoring and adaptive retraining workflows.

Execution

Kafka streaming → online feature engineering → Isolation Forest inference → PSI drift monitoring.

Outcome

Real-time anomaly detection with observability, drift awareness, and operational ML lifecycle management.

Problem

Hospital readmissions increase operational costs and patient risk when high-risk cases are missed.

Approach

Predict calibrated 30-day readmission risk with explainable machine learning.

Execution

Clinical preprocessing → calibrated XGBoost inference → SHAP explanations → drift monitoring and retraining.

Outcome

Explainable risk predictions with monitoring, lifecycle tracking, and clinician-support workflows.

Objective

Control-layer architecture with security filtering, RAG, adversarial multi-agent validation, and self-healing recovery loops.

System

Multi-stage pipeline with input security, RAG, multi-agent validation, and recovery orchestration with full observability.

Results

• Reduced hallucination pass-through to ~5%
• Achieved 100% security block rate

Core Insight

LLMs cannot be trusted as standalone systems — reliability requires external control, validation, and failure-aware orchestration.

Problem

Meetings create information, but decisions and ownership get lost.

Approach

Transform audio into structured decisions and actionable tasks.

Execution

Speech → Text → schema-validated task extraction.

Outcome

Clear decisions, assigned tasks, instant results, reduced manual effort.

Objective

Analyze whether attention heads develop functional roles and quantify their importance using causal ablation.

Method

Built a transformer from scratch and analyzed attention using entropy, positional patterns, and causal ablation.

Key Findings

• Specialization is task-dependent, emerging primarily in early layers
• Head importance is asymmetric; identity heads act as critical bottlenecks

Core Insight

Transformer computation is not uniform — a small subset of heads carries most of the critical behavior.