Reshinth Adithyan

• Led pre-training of StableCode model on a large GPU HPC cluster, which achieved SOTA results. Refer [https://huggingface.co/collections/stabilityai/stable-code-64f9dfb4ebc8a1be0a3f7650]
• Have worked on Pre-training of internal 22B Code Model in TPUs with JAX.
• Large scale data processing and cleaning of pre-training data with pyspark of volumes of text data in TBs.
• Conducted various ablation studies to study nextgen CodeLMs and scaling laws abiding to increase the Reasoning capabilities of CodeLMs.
• Developed methods to increase the reasoning and coding capabilities of base models.

• Responsible for devising DSLs for Clinical Trial Transformation.
• Responsible for research on DSL-based Neural Program Synthesis
• Built a Grounded Neural Search in DSL’s Grammar.
• Program Synthesis embedded in Python for Low Code No Code IDE.

• Responsible for devising, building unsupervised techniques to represent various forms of source codes such as DFG, AST, CFG.
• Responsible for devising methods to compile code using Deep Learning without the use of an explicit Compiler.
• Primarily used, Graph Neural Networks, Transformer based Architectures.
• Expression Evaluation in Conditional Statements for Test Case Generation using Data Flow Graph.

Product - TransformPlus

• Building Graph Neural Network-based Models to represent Naturalness in a Data Flow Graph of a System.
• Naturalness via Hardcoded Values and Variable Names.