ML Engineer · MLDevOps & GenAIOps · Multi-Cloud (Azure · AWS · GCP)

MSc Data Science graduate (Feb 2026) and Associate ML Engineer with 1+ years of industry experience building and maintaining MLDevOps pipelines for automated pharmaceutical document review. I specialize in custom ML model deployment, GenAIOps pipeline engineering, CI/CD automation, RAG systems, multi-agent architectures, and production AI for enterprise solutions.

📧 shivandru@gmail.com · 💼 LinkedIn · 📍 Magdeburg, Germany

Associate Machine Learning Engineer — Quantiphi Analytics Solutions Pvt. Ltd. · Mumbai, India · Aug 2022 – Sept 2023

• Data Engineer (Definity Financial, Apr–Sept 2023) — Migrated legacy Pentaho jobs to Airflow on GCP via Python DAGs; built and validated 17 DAGs for BigQuery warehouse integration.
• Associate MLE (Cerevel Therapeutics, Nov 2022–Apr 2023) — Built and maintained Cloud Functions running MLDevOps pipelines for automated medical-document review from an eTMF server; shipped 4 custom classification pipelines (Financial Disclosures 3K PDFs, PPD Reports 1.5K, Hypercare SOWs 2K, FDA 1572 1K); trained Vertex AI AutoML models with custom Document AI parsers.
• MLE Intern (Unum Insurance, Aug–Nov 2022) — Built AutoML pipeline tagging 1,000+ medical underwriting documents via Cloud Functions.

Latency-Budgeted Onset and Note Detection (GCP) (Nov 2025 – Feb 2026) — real-time music transcription closing the distribution-shift gap between offline training and memory-constrained deployment, extending Hu et al. (ISMIR 2025) with a matched-training methodology.

• Trained 42 models (3 architectures × 5 latency budgets × 3 conditions) on Vertex AI, orchestrating parallel jobs across T4/L4/A100 GPUs with Docker and automated GCS result upload
• Solved a deployment distribution-shift problem (models trained on 3–10s sequences degrading under 370–670ms real-world memory constraints); matched training outperforms truncated inference across all 15 architecture-budget combinations (+0.87 pp avg F1)
• Discovered via controlled ablation that a stepped geometric curriculum (500→250→125→62→target frames) is essential — removing it drops performance by 2.67 pp

→ LatBOND-Framework · Python · PyTorch · Vertex AI

Production-pattern implementations demonstrating enterprise MLOps, GenAIOps, and edge deployment.

Enterprise document search retrieving answers from technical manuals with zero PII exposure under GDPR.

• End-to-end GDPR-compliant RAG pipeline (ingestion → PII detection → hybrid retrieval → grounded generation)
• Custom German PII recognizers (IBAN checksum, Steuer-ID check digit) at 0.95+ recall; RAGAS eval (groundedness 0.80+, faithfulness 0.85+) as a CI promotion gate
• Bicep IaC (ML workspace, AI Search, Container Apps, Key Vault) + GitHub Actions with 75%+ coverage target

Azure OpenAI · AI Search · Container Apps · Presidio · FastAPI · MLflow · Docker · Bicep

Autonomous agents detecting supply chain disruptions and proposing cost-optimised mitigations with a SAP-compatible data layer.

• LangGraph multi-agent system (conditional routing, parallel execution, cycle detection) with Cosmos DB checkpointing for full decision replay
• Bayesian risk scorer + OR-Tools LP optimizer (15% cost savings vs naive allocation), 4-tier progressive constraint relaxation, HITL escalation
• SAP S/4HANA-compatible mock (MM/SD, 50 suppliers, 200 materials, 500 POs) with Repository pattern for zero-code swap to live OData v4; Hypothesis property tests (200+ examples)

Azure OpenAI · Cosmos DB · Event Grid · LangGraph · OR-Tools · Prophet · FastAPI · MLflow · Bicep

End-to-end pipeline compressing pre-trained automotive models for edge hardware with ISO 26262 safety gates.

• QAT engine with per-layer sensitivity analysis and mixed-precision (INT8/FP16/FP32) — 4× compression at <1% accuracy loss; ONNX export with target-runtime validation
• 3-detector drift ensemble (PSI + MMD + ADWIN) with weighted voting and temporal persistence filter — auto-rollback under 60s
• BFS requirements-traceability graph enforcing 100% ISO 26262 coverage (deployment hard-blocked on gaps); canary rollout (5/50/100%) via Terraform

SageMaker · Neo · Model Registry · Model Monitor · IoT Greengrass · ONNX · TensorRT · Terraform

Cloud: Azure (OpenAI, AI Search, Container Apps, Durable Functions, Event Grid, Cosmos DB, Microsoft Foundry) · AWS (SageMaker, Model Monitor, CodePipeline, IoT Greengrass) · GCP (Vertex AI, BigQuery, Document AI, Cloud Functions)

AI/ML: RAG (hybrid BM25+dense, RRF, RAGAS) · Multi-Agent (LangGraph) · LLMs (GPT-4o, prompt flow) · Vector DBs (AI Search, Qdrant) · PyTorch · TensorFlow · SpaCy · ONNX · AutoML · Bayesian Risk Scoring · OR-Tools LP

MLOps / IaC: MLflow · SageMaker Pipelines · Airflow · Model Registry · Drift Detection (PSI, MMD, ADWIN) · CI/CD (GitHub Actions, CodePipeline) · Docker · Bicep · Terraform

Enterprise: GDPR/Datenschutz (Art. 5/6/17/25/30/35, Presidio) · ISO 26262 · SAP S/4HANA (MM/SD, OData v4) · OWASP LLM Top 10 · Human-in-the-Loop

Languages: Python (PyTorch, FastAPI, Pydantic, asyncio, Hypothesis) · SQL · JavaScript · Bash

MSc Data Science — University of Europe for Applied Sciences, Potsdam · Sept 2024 – Feb 2026

BE Computer Engineering — Lokmanya Tilak College of Engineering, Mumbai University · 2012 – 2018

• Microsoft AI-300 — Operationalizing ML and GenAI Solutions (MLOps Engineer Associate) · Pursuing
• Microsoft AI-103 — Azure AI App and Agent Developer Associate · Pursuing

English (C1, IELTS 7.0) · German (A2, B1 exam Jul–Aug 2026) · Malayalam (Native) · Hindi (Fluent) · Marathi (Fluent)