Yuvaraj Sriramoju

Corebridge Financial

Generative AI Engineer · Sep 2024 — Present · Dallas, TX

Architected the agentic AI platform behind underwriting, policy analysis, and compliance — specialised agents orchestrated over enterprise APIs, vector stores, and a Neo4j graph.

• Production agentic platform on MCP, LangGraph and LangChain, coordinating specialised agents against enterprise APIs and vector databases
• Enterprise RAG with hybrid retrieval, reranking and vector memory — answer relevance up 28%
• Neo4j knowledge graphs wired into retrieval for relationship-aware reasoning — policy analysis accuracy up 30%
• Fine-tuned GPT, BERT, T5 and diffusion models for document summarisation — 32% accuracy gain, 45% less manual review
• Prometheus and Grafana telemetry on LLM latency, token spend and inference drift — downtime down 35%
• Model cards, data lineage and governance reports for GDPR, CCPA and FINRA audit readiness

LangGraph · MCP · Neo4j · FastAPI · Azure ML · MLflow

Tenet Health Care

Machine Learning Engineer · Oct 2023 — Aug 2024 · Dallas, TX

Clinical ML across EHR, claims and imaging — readmission risk, radiology triage, and the MLOps to keep it all HIPAA-clean.

• Readmission-risk and length-of-stay models on EHR and claims data — risk stratification accuracy up 28%
• Multimodal CV pipeline with CLIP, BLIP and PyTorch for X-ray anomaly detection — diagnostic accuracy up 19%
• BERT-based clinical NLP extracting diagnoses, medications and symptoms from physician notes at 87% F1
• Reinforcement-learning treatment prototype improving decision alignment with clinicians by 18%
• MLOps on SageMaker and Kubernetes with automated retraining — 99.5% uptime, incident response 40% faster
• Encryption, de-identification and SHAP/LIME explainability for HIPAA and PHI compliance audits

PyTorch · CLIP / BLIP · SageMaker · Kubernetes · PySpark · FHIR

Kotak Mahindra Bank

Data Scientist · Mar 2022 — Jul 2023 · Bengaluru, India

Credit, fraud and forecasting models serving real-time decisions across lending and payments.

• Credit scoring and repayment-risk models on XGBoost and LightGBM — prediction accuracy up 22%
• Fraud false positives down 15% through supervised learning and feature-importance analysis
• LSTM and Prophet forecasting for loan demand and liquidity — 25% more accurate capital allocation
• Recommendation engine matching credit and loan products to transaction behaviour and credit profiles
• Risk and fraud models deployed as REST microservices for real-time credit and payments decisions
• SHAP, LIME and lineage documentation for RBI, GDPR and internal audit compliance

XGBoost · LightGBM · Prophet · MLflow · Kubeflow · REST

SG Analytics

Data Scientist · Aug 2020 — Feb 2022 · Bengaluru, India

Where the pipelines started — ETL, forecasting and BI for finance and healthcare clients.

• ETL across multi-source client datasets in SQL, Python and Snowflake — refresh latency down 32%
• ARIMA, regression and churn models for market forecasting — predictive accuracy up 21%
• Validation and reconciliation automation in Python and Pandas — 45% less manual QA
• Power BI and Tableau dashboards cutting reporting turnaround from three days to under one
• Data lineage, versioning and feature-store frameworks for reproducible, auditable pipelines

Snowflake · Python · SQL · Power BI · Tableau · Pandas

Cartographer

MCP server · In Development

An MCP server that exposes a Neo4j graph as typed tools — fetch neighbours, find shortest path, extract a subgraph, resolve an entity. Any MCP client gains relationship-aware reasoning without writing a line of Cypher.

Python · MCP SDK · Neo4j · Pydantic

Crucible

Evaluation harness · Planned

Generates adversarial questions from your own corpus — multi-hop, negation, near-miss distractors — then scores retrieval recall@k, answer faithfulness and hallucination rate. Runs in CI so a regression fails the build.

Python · FAISS · pytest · GitHub Actions

Radiograph

Multimodal research · In Development

A vision encoder feeding an LLM decoder that emits a structured report where every finding carries the image region that produced it. A radiologist can check any sentence against the pixels behind it.

PyTorch · ViT · CLIP · MIMIC-CXR

Vault

Private retrieval · In progress

Fully air-gapped RAG. Quantised embeddings and a local model index a document set on a laptop with zero network egress, plus an append-only audit log of every retrieval performed.

llama.cpp · ONNX Runtime · FAISS · FastAPI

Driftwatch

Observability · In progress

A sidecar that samples production traffic, embeds prompts and responses, and alerts when either distribution shifts. Catches prompt drift, response drift and retrieval decay before a user reports them.

Python · Prometheus · Grafana · sentence-transformers

Consilium

Multi-agent environment · Prototype

An environment where agents hold private information and incompatible goals, must negotiate through tool calls, and get scored on both outcome and honesty. Every run produces a replayable trace.

LangGraph · Python · asyncio