Samsushma A

Agentic AI / Data Engineer at New Math Data, LLC

Built enterprise legal AI platform for automated legal document analysis, patent claim intelligence, risk classification, and compliance workflows deployed for attorneys with full auditability and governance. Developed production-grade agentic patentability research using AWS Step Functions, distributed workflows, and Amazon Bedrock across 100 concurrent invocations per workflow run. Implemented multimodal disclosure pipeline ingesting 7+ inventor file formats (PDF, DOCX, audio, video) routed through format-specific extraction and Bedrock structured JSON generation. Engineered RAG-based semantic prior art search using OpenSearch vector search with reranking, MCP tool use, and async fan-out for N+1 parallel query execution. Implemented a 5-phase transformer-based LLM reasoning pipeline with Bedrock prompt caching, reranking, and state management, achieving high prompt-cache hit rate and risk assessments with evidence chains; also shipped LLM pipelines using Lambda, Docker, CI/CD, and Te

Agentic AI/ Data Engineer at New Math Data, LLC

Built an enterprise legal/AI patent platform automating legal document analysis, patent claim intelligence, risk classification, and compliance workflows with end-to-end auditability. Developed production-grade agentic patent research using AWS Step Functions, distributed mapping/Lambda orchestration, and Amazon Bedrock across up to 100 concurrent invocations per workflow run. Assembled multimodal disclosure pipelines ingesting 7+ file types (PDF/DOCX/audio/video) via Bedrock and AWS Transcribe/Nova Pro with schema-validated JSON outputs. Implemented RAG semantic prior art search using OpenSearch vector search, embedding retrieval, and cross-encoder reranking with MCP tool use. Engineered a 5-phase transformer/LLM reasoning pipeline enforcing structured output schema and guardrails to produce cited-evidence risk assessments with no hallucinated citations.

AI/ML Engineer at New Math Data, LLC

Built agentic AI patent-ability and data pipelines, integrating real-time portfolio insights, risk assessment, and compliance monitoring for 500+ investment professionals. Engineered multi-agent orchestration with LangGraph and MCP using LLaMA-3/4 for research, compliance, and reporting. Delivered an end-to-end data stack and governance framework with prompt-engineering, RAG retrieval, and semantic reasoning across 13 schema collection types, enabling scalable, auditable AI-driven narratives.

Agentic AI Narrative Designer at AgentCore Runtime / Arcweave

Architected and shipped an autonomous narrative-generation platform on AWS Agent Core Runtime and Amazon Bedrock, including a 7-step inference pipeline and automated content generation with zero manual editing. Led end-to-end design of Arcweave-compatible JSON schemas and fidelity controls for multi-scheme data collection, enabling consistent multi-board generation and keeping context through prompting. Implemented dedicated asset generation workflows (images, video, scene art) and automated validation to ensure high-quality outputs with reduced human intervention. Established a human-in-the-loop workflow feeding validated prompts back into the system for targeted story graph corrections, significantly reducing manual content creation effort.

Agentic AI/Data Engineer at New Math Data, LLC

Agentic AI/Data Engineer at New Math Data, building an AI-enabled platform for patentability research and multi-agent orchestration on AWS Step Functions, enabling scalable, automated IP research with low hallucinations and end-to-end governance.

GenAI/ML Engineer at Fidelity Investments

Automated Chief Investment Officer workflows integrating portfolio optimization, real-time market intelligence, and compliance monitoring serving 500+ investment professionals. Built predictive portfolio allocation models using XGBoost and scikit-learn and improved risk-adjusted allocation strategies. Orchestrated data pipelines using Apache Kafka, AWS Glue, and Athena to structure event streams for downstream GenAI agents including time-series feature windows and cross-sector risk indicators. Executed multi-agent orchestration with LangGraph and MCP using Llama for research, compliance, and reporting; reduced manual memory creation time. Fine-tuned LLMs using Hugging Face Transformers with QLoRA/PEFT for financial summarization and document classification while reducing GPU memory; built prompt engineering systems and deployed RAG indexing for financial documents/filings using FAI/FAISS and Azure vector retrieval. Implemented MLflow support for MLOps/LLMOps, retraining and drift detec

GenAI/ ML Engineer at Fidelity Investments

Automated Chief Investment Officer workflows by integrating portfolio optimization, real-time market intelligence, and compliance monitoring serving 500+ investment professionals. Built predictive portfolio allocation models using XGBoost and scikit-learn, improving risk-adjusted allocation strategies by 12% in live trading. Orchestrated data pipelines with Apache Kafka, AWS Glue, and Athena at 99.8% uptime to stream market feeds and regulatory updates into structured event streams for downstream GenAI agents, including feature engineering for time-series windows and cross-sector risk indicators. Implemented multi-agent orchestration with LangGraph and MCP using Llama 3 8B for research/compliance/reporting and reduced manual memory creation time by 50%. Fine-tuned LLMs (FinGPT, Palmyra, Gemini) using Hugging Face Transformers with QLoRA/PEFT to reduce GPU memory by 60% while improving financial summarization and document classification; implemented RAG with indexing and hybrid rerankin

AI/ML Engineer at FIDELITY INVESTMENTS

Built GenAI pipeline automating Chief Investment Office workflows, integrating portfolio optimization, real-time market intelligence, and compliance monitoring, serving 500+ investment professionals. Implemented multi-agent orchestration with LangGraph and MCP using LLaMA 3 8B for research, compliance, and reporting; applied semantic routing and task-specific prompts to cut manual memo creation time by 50%. Optimized FinGPT/Palmyra with QLoRA and PEFT for financial summarization and document classification, reducing GPU memory usage by 60%. Established a RAG pipeline leveraging FAISS and Redis for indexing earnings reports and regulatory filings, with production indexing on Azure Vector DB using HNSW + hybrid BM25/semantic reranking, achieving 35% improvements in retrieval relevance.

AI/ML Engineer at Capgemini

Delivered a fraud detection system using PyTorch and deep neural networks, reducing false positives by 31% and improving anomaly detection in high-volume transactional environments. Calibrated classification models to address severe class imbalance via SMOTE, boosting recall across multiple financial institutions. Built NLP pipelines with Hugging Face Transformers (BERT) and custom tokenization to extract risk entities, intent signals, and customer sentiment from KYC documents; enabled real-time predictions via FastAPI with Lambda, achieving sub-200 ms latency. Orchestrated end-to-end workflows with Airflow and Spark windowed aggregations; supported model retraining and drift monitoring with MLflow.

AI/ML Engineer at Capgemini

Delivered an autonomous research platform for LLM-based patent analysis; built multi-modal, multi-step RAG pipelines with AWS Bedrock and OpenSearch; implemented a 5-phase LLM reasoning pipeline with prompt chaining and semantic ranking, achieving significant reduction in manual IP research effort.