Abhishek Kanniganti

Led a team of 7 AI Engineers in delivering AI-driven solutions across diverse industries. Experienced in the full lifecycle of machine learning projects—from use-case identification and model development to deployment and productionization.

Projects:

• Supply Chain Modernization for Leading Laptop Manufacturer: Led a team of 10 engineers to develop an automated low-code LLMOps and MLOps platform, streamlining the deployment of AI and GenAI solutions. Integrated frameworks such as Langfuse and MLflow, and implemented CI/CD pipelines with Kubernetes-based deployments, to accelerate onboarding. Successfully reduced the average productionization time for supply chain AI use cases from 2 months to 1 month. Incorporated robust features, including Automated Agent Evaluation, Drift Monitoring, and LLM Observability, to enhance system reliability and maintainability.
• First Draft Response Automation: Developed an advanced RAG and NL2SQL-based response generation asset, leveraging large language models to reduce the efforts required by analysts for writing RFP responses. Additionally, I created a robust pipeline to extract data from PDFs, generate outlines, win themes, and past success stories for a given RFP through LLMs like GPT-4 and Mistral, and perform embedding searches over the knowledge base of unstructured PDFs and a structured database. Created a comprehensive evaluation strategy for evaluating and benchmarking various components of the solution. The application has been productionized with dashboards for token usage, GPU utilization, tracing, and logging, and is currently used by three teams. Its knowledge base has been scaled to over 1,000 documents.
• Readmission Prediction: Finetuned LLama 3.1 70B-based model to predict patient readmission based on clinical notes for a large hospital in New York. Achieved 75% accuracy on the test data and deployed the model on VLLM after performing various inference optimizations using the Nvidia Nemo framework.
• Converter Failure Prediction for Wind Turbines: Led a team of three analysts to develop a predictive maintenance solution to identify converter failures in advance. Analyzed data from 17 sensors and extracted key patterns, resulting in converter failure. Developed explainable models to predict failures in converters in advance. As of the first phase, we achieved a 30% accuracy lift in detecting converter failures and a 60% reduction in false alarms compared to their existing systems.
• Hot Socket Prediction: Developed an ML model using Auto-Encoder and XGBoost to predict high-temperature events, or hot-socket events, in smart meters in advance, and help businesses prevent emergencies resulting from them. Built data processing pipelines leveraging various distributed frameworks like Pyspark, Dask, Rapids to efficiently process 120 TB of sensor data on a GPU cluster. The model is deployed in production and evaluates the sensor data of 4.3 million smart meters in a city and gives the business a list of meters which are going to undergo hot-sockets. The event is a rare event, occurs in about 250 meters out of 4.3 million meters over the period of one year. The model captured 80% of hot-sockets and has a hit-rate of 55%.

• Worked on developing IMIBot’s primary NLP engine and other internal tools by enhancing an open-source library to transform IMIBot from code-based to GUI-driven. This has led to reduction in time taken to build a bot by 80 %.
• Enhanced the IMIBot platform to handle more than 100K messages per year, and it has improved the accuracy of bots by 10% compared to a rule-based engine.