MLOps & Machine Learning
From raw data to production models — feature pipelines, training, eval harnesses, containerized serving, and live monitoring across four production environments.
Intuitive
ML Engineer (Co-Op) • Sunnyvale, CA • Jun 2025 – Mar 2026
Framed manual support as a data problem. Ran EDA in Python & Spark SQL to characterize the ticket landscape, then applied BERTopic clustering and embedding retrieval over ServiceNow & chat text to surface high-volume, automatable workflows.
Discovery EDA
Python/Spark SQL analysis across ServiceNow & live chat logs to quantify automation opportunity surface.
BERTopic Clustering
Unsupervised topic modeling over support text to group and rank ticket types by automation ROI.
Embedding Retrieval
Dense embedding retrieval over historical tickets to surface semantically similar resolved cases for grounding.
Azure Databricks & PySpark
Data Scientist Intern • Bangalore, India • Jun–Aug 2023
Built ETL & PySpark feature pipelines on Azure Databricks over Delta Lake, orchestrated as scheduled Databricks Workflows jobs with Unity Catalog governance over 10+ years of time-series data.
Train. Track. Reproduce.
Every model that ships starts with rigorous experimentation — classical ML baselines, gradient-boosted trees, and transformer fine-tuning — all tracked for full reproducibility.
Intuitive
Ticket Triage & Deflection Models
Logistic regression, gradient-boosted trees, HuggingFace transformers
- Trained logistic regression & gradient-boosted tree classifiers for ticket triage and deflection scoring.
- Fine-tuned HuggingFace/PyTorch transformers on domain-specific support text for intent classification.
- MLflow tracking — params, metrics, artifacts, and model registry across every experiment run.
UPL
Demand Forecasting Models
XGBoost, LightGBM, ARIMA with time-series CV & SHAP
- Trained XGBoost, LightGBM & ARIMA models on 10+ years of agricultural demand time-series.
- Validated with time-series cross-validation to prevent data leakage across temporal folds.
- Applied SHAP for feature importance and explainability — communicating model logic to stakeholders.
BM25 + Dense HNSW + Cross-Encoder Reranking
Engineered a hybrid search pipeline combining BM25 lexical scoring with dense embedding retrieval (HNSW index) to capture both exact keyword matches and semantic meaning. Added cross-encoder reranking to lift grounding accuracy and relevance for the RAG pipeline serving production agents.
Also built agentic AI loops in LangChain with guardrails and safety controls against prompt injection and data leakage, deployed in production at Intuitive.
Lexical Retrieval (BM25)
Sparse keyword scoring captures exact terminology — critical for domain-specific ticket language and product names.
Semantic Retrieval (HNSW)
Dense embedding index (HNSW) retrieves semantically similar docs even when phrasing differs — covers paraphrase and intent variation.
Cross-Encoder Reranking
A cross-encoder scores query-document pairs jointly, reranking the merged candidate set for maximum grounding accuracy before generation.
Safety Guardrails
Agentic loop guardrails enforce prompt injection defense and data-leakage controls for safe, compliant output in enterprise settings.
Eval & Gated Release
Every release is gated on regression thresholds. Whether LLM-as-judge scoring or statistical backtesting, rigorous evaluation is a first-class citizen.
Intuitive — LLM Systems
Eval Harness & LLM-as-Judge
- Built eval harnesses & benchmarks scoring grounding accuracy and retrieval quality.
- LLM-as-judge evaluators automated grading at scale across retrieval and generation pipelines.
- Every release gated on regression thresholds — no rollout without passing the eval suite.
ION Group — Regulated Finance
Quantitative Model Validation
- Independently recomputed VaR (historical & Monte Carlo) as a golden reference, benchmarking a regulated production risk engine.
- Backtested VaR against realized P&L using Kupiec-style exception tests (SciPy/statsmodels).
- Shipped automated QA regression reports via CI/CD, replacing manual pre-release checks.
Dronaid — Edge ML
Active Learning & CV Model Evaluation
Curated large-scale drone image datasets with active-learning sampling to select the most informative frames. Evaluated CV models on precision, recall, mAP & error analysis before promoting to on-device deployment.
Data Quality Gate
Active-learning loop selects only the most informative frames, avoiding redundant annotation.
Imbalance Handling
EDA-driven sampling strategy to correct class imbalance on rare aerial categories.
mAP Threshold Gate
Models only promote to deployment when mAP meets minimum thresholds on holdout splits.
Error Analysis
Systematic analysis of false positives/negatives to guide next training iteration.
MLflow Registry to Containerized Endpoints
Models trained at Intuitive were promoted through the MLflow model registry into containerized REST endpoints. CI/CD pipelines enforced versioning, reproducible builds, and rollback strategies for safe production releases.
At Dronaid, the deployment target was an on-device C++ runtime on embedded Linux — models were optimized with TensorRT + CUDA for real-time, latency-bound edge inference with strict memory constraints.
See C++ & CUDA edge deployment showcaseMLflow Model Registry
Staged promotion: Staging to Production with lineage tracking, artifact storage, and version pinning across all model types.
Containerized REST Serving
Models served as Docker-containerized REST endpoints with environment isolation and reproducible runtime dependencies.
CI/CD Pipelines & Rollback
Automated build, test, and deploy pipelines with versioning discipline and rollback strategies for zero-downtime safe releases.
Edge Inference (Dronaid)
TensorRT + CUDA optimized inference deployed in C++ on embedded Linux for real-time drone perception within tight latency and memory budgets.
Own Production End-to-End
After deployment, I owned production support at Intuitive: monitoring for data drift, latency SLAs, and anomalies. Ran rigorous A/B experiments with hypothesis testing to validate lift before full rollout — tracked in Power BI dashboards.
Drift & Latency Monitoring
Production dashboards tracking data drift, model latency, and SLA compliance — alerting on regressions before users notice.
A/B Testing & Hypothesis Validation
Ran controlled A/B experiments with t-tests and proportion tests to statistically validate performance improvements before full rollout.
65% Support Volume Automated
The compound result: discovery to training to eval to deployment to monitoring closed the loop, automating 65% of support volume end-to-end.
Tools That Have Shipped to Prod
Training & Experimentation
- PyTorch
- HuggingFace Transformers
- XGBoost
- LightGBM
- Scikit-learn
- ARIMA
Tracking & Registry
- MLflow Tracking + Registry
- Experiment Params & Artifacts
- Reproducible Builds
- Version Pinning
Data & Feature Engineering
- PySpark / Spark SQL
- Azure Databricks
- Delta Lake
- BERTopic
- Pandas / NumPy
- Unity Catalog
Retrieval & RAG
- BM25 Lexical Search
- HNSW Dense Retrieval
- Cross-Encoder Reranking
- LangChain
- Embedding Models
Deployment & Serving
- MLflow Model Registry
- Docker Containers
- REST Endpoints
- TensorRT + CUDA
- CI/CD Pipelines
- Rollback Strategies
Eval & Monitoring
- LLM-as-Judge
- Regression Thresholds
- A/B Testing
- Kupiec Backtesting
- Drift Monitoring
- Power BI Dashboards