Malhar Thombare

Experience

• 

  CereLabs

  Jun 2016 - Jul 2021

  Intelligent System for Corporate knowledge acquisition and querying. (NLP, Graph Databases: Neo4j, OrientDB)Designed and developed a platform to extract information from unstructured corporate data (text, tables), store it as retrievable knowledge, and perform reasoning and inferencing over the knowledge to answer questions.Developed a programming paradigm to build custom knowledge extraction and inferencing pipelines.As part of pipeline builder, implemented image-to-text (YOLO, tesseract), text cleaning, text-to-knowledge graph (OpenIE libraries), reasoning over knowledge graph, and automated common-sense acquisition modules (OpenIE libraries).Served these modules in a service-oriented architecture with an ability to integrate user-built modules seamlessly. Show less

  • A.I. Researcher

    Jul 2017 - Jul 2021

  • Research and Development Intern

    Jun 2016 - Jul 2017
• 

  University of California, Riverside

  Mar 2022 - May 2023

  Emotion-guided difficulty adjustment of a videogame which helps memory-affected patients.Based on small existing data, developed a simulator of player’s performance and facial expressions (Stochastic Processes).Successfully demonstrated Meta Reinforcement Learning (A2C)’s application to solve the problem.Collaborated with the facial expression recognition team and the psychology department’s team. Improving Bayesian Optimization for Quantum Material Control (Master’s Thesis).Developed a 10x faster variant of BO to find an electromagnetic signal that would control the quantum state of an atom (in the NIC-CAGE simulator).Designed an experimentation framework to test multiple BO variations on tens of thousands of quantum system simulations on a compute cluster.

  • Graduate Student Researcher

    Jun 2022 - May 2023

  • Graduate Student Researcher

    Mar 2022 - Mar 2023
• 

  Beyond Limits

  May 2023 - Jun 2024

  AI Engineer

  Built a document-to-code application to encode industrial knowledge into a company’s proprietary programming language.Designed the product, in collaboration with the product team, to satisfy market requirements with company technology.Designed and built an automatic syntactic and semantic check for quality assessment of the AI-generated code.Performed prompt engineering including quality and speed assessment on a range of LLMs/Foundation models:Closed source: GPT (3.5, 4, 4 turbo), Claude 3 (Haiku, Sonnet, Opus).Open source: Mistral 7B, Mixtral 8x7B, Llama 2, Mixtral-dolphin.Developed a technique to generate AI-written programs a lot longer than 4000 tokens (current limitation of LLM APIs)Further improved code generation with verification and automatic correction by post-processing AI-generated code. Show less