Alekhya Vuppala

Senior AI/ML Engineer building production Generative AI and agentic AI systems end to end: LLM-driven multi-step agent workflows, RAG pipelines, graph-based recommendation engines, and LLM evaluation frameworks.

Experience across financial services and legal-tech domains. Owns the full stack from schema design, prompt strategy, embeddings and vector retrieval through FastAPI microservices and cloud deployment on GCP and AWS. Experienced working directly with clients to translate business requirements into robust AI architecture, with a focus on output reliability, validation, and responsible AI practices.

Senior AI Engineer (Freelance) - Prudentix

(2026-05)

• Architected EventFlows, a DAG-based agentic workflow automation platform that generates and executes multi-step task workflows from natural language, supporting 15 node types including an "Apps" node that integrates 25+ MCP connectors for tool use.
• Designed the JSON Schema spec, Pydantic v2 validation models, and LLM system prompt driving automated, structured workflow generation.
• Implemented a pgvector-based RAG pipeline for workflow retrieval and generation with Gemini, using context caching for the node catalog and system prompt.
• Built the FastAPI microservice layer (REST APIs) and deployed it on GCP Cloud Run; resolved CodeQL CI failures to keep the CI/CD pipeline reliable.
• Built a DeepEval-based golden test set framework to evaluate generation quality, with schema-validation gating ahead of LLM-as-judge scoring to catch malformed or unreliable outputs.

Senior ML Engineer - Micro1

(2023-08 - 2024-10)

• Designed and deployed a graph-based recommendation engine using Neo4j and Leiden community detection, shipping it to a production pilot with 1,000 users to validate recommendation quality and system stability ahead of broader rollout.
• Implemented RLHF-inspired feedback loops to incorporate live user interactions into iterative model retraining, closing the loop between real usage signals and recommendation relevance.
• Built a simulated agent-based environment to stress-test the recommendation system beyond the pilot user base, surfacing edge cases and validating behavior under higher load.
• Worked directly with the client while the application was being built, gathering requirements and translating business needs into recommendation-system design and deployment decisions.
• Owned end-to-end ML architecture from data processing to deployment (PostgreSQL, FastAPI, AWS EC2/Lambda), designing for horizontal scalability from pilot to production-scale traffic.

ML Engineer (Contract) - Bluecopa

(2023-02 - 2023-05)

• Built an LLM-powered internal Q&A chatbot using GPT-3.5, Llama, LangChain, and Pinecone, letting users ingest company documents and query them in natural language, replacing a manual review process that previously took days across thousands of documents.
• Implemented a RAG architecture with embedding-based similarity search over the ingested corpus, paired with an ML-based outlier detection system to flag anomalies in financial transactions within the same pipeline.

ML Engineer - Contract Wrangler

(2017-08 - 2020-08)

• Built an NLP date extraction and classification system for contracts using regex parsing, a scikit-learn SVM classifier, and spaCy for relative date resolution, achieving a 94% F-score; ran automatically on every new document upload with results persisted to MySQL.
• Built an information extraction pipeline using spaCy and regex to pull key contract metadata: titles, governing law, and party names.
• Built a clause classification pipeline from ~10K scraped contract clauses; applied cosine similarity deduplication on sentence embeddings (~30% reduction), then clustered the remainder into ~100 clause archetypes (K-Means/HDBSCAN), reducing manual labeling effort by ~70%.

Software Engineer - FactSet Systems

(2015-03 - 2016-12)

• Built classifiers for financial press release articles and financial table categorization using Weka (Java) and Apache Spark ML.
• Built a generic document vectorization pipeline (word2vec, TF-IDF) to support downstream classification tasks.

MS - Computer Science - University of Southern California (2015)

B.Tech - Computer Science - SRM University (2013)