AI Engineer · 9+ years
Akhilanand Benkal Venkanna
Summary
Lead ML Engineer with 9+ years building production generative AI, machine learning, NLP, and data systems on AWS and Databricks. Specializes in AI agents, retrieval-augmented generation, text-to-SQL, and LLM evals, including an agentic LangGraph assistant used by 1,400 field reps and 50 regional managers.
Experience
Johnson & Johnson
AI Sales Assistant: agentic LLM orchestration
Python, LangGraph, FastAPI, Pydantic, Claude (Opus, Sonnet, Haiku) on AWS Bedrock, Textract, Amazon OpenSearch, PostgreSQL, RAGAS, Databricks MLflow
- Shipped an agentic LangGraph assistant for 1,400 field reps and 50 regional managers, routing questions to text-to-SQL, hybrid RAG over Textract-extracted PDFs, SKU lookup, and account search; cut time-to-answer by ~30%.
- Orchestrated three Claude models on AWS Bedrock: Opus for accuracy-critical routing and multi-step tool calls, Sonnet for answers, and Haiku for query rewriting, serving ~2,000 queries a day through FastAPI.
- Improved routing accuracy from 80% to 95%+ through iterative error analysis; added LangGraph retry loops for failed SQL and clarifying questions when queries returned 100+ rows.
- Lifted RAGAS faithfulness from 0.75 to 0.9+ with top-k tuning, reranking, and source-only answering; used GPT judges for groundedness and guardrail adherence and traced live traffic in Databricks MLflow.
Pricing engine: team lead
Python, scikit-learn, price elasticity and optimization, customer segmentation, PySpark, Airflow
- Led a team of 6 delivering a pricing engine that recommends line-item discounts from price elasticity and optimization, realizing $4M a year by curbing excess discounts on new products; built its customer segmentation models.
Product classification and SKU matching
Python, XGBoost, Random Forest, TF-IDF, PubMedBERT, Sentence Transformers, FDA GUDID
- Designed a three-stage XGBoost and Random Forest pipeline classifying ~400K products from customer files and FDA GUDID, filtering out non-devices and mapping each to a J&J platform and segment at 95% and 92% accuracy.
- Cut analyst mapping time by 40% with a one-click shortlist of the 20 closest J&J SKUs, using TF-IDF to narrow candidates to 100 and a fine-tuned PubMedBERT model to rank them at 76% top-20 accuracy.
Johnson & Johnson
- Built the daily Airflow and PySpark pipelines on AWS for a COVID-19 dashboard that forecast sales at the zip-code level from projected case spread and hospitalization trends.
- Migrated local data science workloads to EMR and SageMaker, rewriting complex joins in PySpark and tuning Spark jobs to cut runtimes by 60%.
Monsanto
- Engineered streaming ETL in Scala, Kafka, and Akka for the Customer360 asset platform, consuming 20+ source systems and processing millions of records.
- Developed RPC APIs in Go and deployed services as Docker containers on Monsanto's internal Kubernetes cluster.
Colaberry
- Integrated Oracle, Teradata, SAP, and Salesforce into Monsanto's Customer360 platform through Kafka Connect and Scala Akka Streams ingestion services, containerized with Docker.
Skills
LLM
LangGraph · LangChain · prompt engineering · vector search · Claude · AWS Bedrock · OpenAI GPT · RAG · embeddings · rerankers · tool calling · Pydantic · guardrails · text-to-SQL
ML and evaluation
LLM-as-judge · RAGAS · Databricks · MLflow · PyTorch · Sentence Transformers · embedding fine-tuning · XGBoost · NER · OCR
Data and cloud
Spark · PySpark · Airflow · Kafka · AWS SageMaker · EMR · OpenSearch · Textract · FastAPI · Docker · Kubernetes · PostgreSQL
Languages
Python · SQL · Scala · Go
Education
Illinois Institute of Technology
Master of Science, Computer Science · Chicago, IL
Aug 2015 – May 2017