Machine Learning Engineer | ML Systems • Time-Series • Reliability
I build production-oriented ML systems where labels are missing, noise is high, and failures are expensive. My focus is on how ML behaves under real-world constraints, not just model accuracy.
- Label-free & weakly supervised ML problems
- Time-series modeling and drift detection
- Full ML Lifecycle: Preprocessing, Training, Tuning, Validation & Testing
- Offline learning + deterministic online systems
- Failure modes, alert fatigue, and explainability
I prefer statistical ML and explicit objectives over opaque black-box models when reliability matters.
Hybrid ML System for Streaming Regime Shift Detection
%20Deterministic-success?style=flat-square){kind=icon}
BLACKICE detects persistent behavioral drift in infrastructure metrics using a hybrid ML architecture.
| ML Problem (Constraints) | ML Approach (Solution) |
|---|---|
| No labeled data | Streaming statistical baselines (Welford) |
| Highly noisy, bursty signals | Offline optimization of decision boundaries |
| False positives > delays | Custom SRE-weighted loss function |
| Black-box opacity | Persistence-aware detection (not point anomalies) |
Impact: ~80–90% noise filtered, <1% false positives, O(1) memory, tested on 8GB+ production data.
🔗 Repo: https://github.com/Mihirmaru22/blackice
Real-time forest fire risk assessment API
- Precision Modeling: Ridge Regression pipeline with automated feature scaling.
- Production Ready: Serialized Joblib model served via RESTful Flask interface.
- Deployment: Container-friendly structure for AWS/Render.
Tech Stack
Currently Learning
