Skip to main content
Harshad Shinde

Harshad Shinde

Data Scientist

Talent
PuneMember since August 14, 2026
PythonC++SQLPyTorchHuggingFace TransformersTRLDeepSpeedLoRAScikit-LearnNumPyPandasSFTGRPODDPmulti-GPU distributed trainingRLHFreward modelingvLLMFastAPIGCPVultrSnowflakeDICOMMySQLGitJupyterGoogle ColabVS CodeLinuxLINQ
·

Hire this person

Send a job offer directly to this candidate

PythonC++SQLPyTorchHuggingFace Transformers

About

M.Tech (IIT Kharagpur) data scientist who owns the full ML lifecycle — from problem definition and framing the ML approach, through data curation and dataset design, model fine-tuning, evaluation, and deployment. Specializing in fine-tuning and serving large vision-language models for enterprise workflow automation.

Experience

Data Scientist - NeuroLogic AI Systems Pvt. Ltd.

(2025-01)

Enterprise workflow automation with vision-language models

  • Fine-tuned Qwen2.5-VL-32B-Instruct (LoRA) to both predict the next action and understand enterprise claim-filing workflows; handled checkpoint consolidation and adapter merging for inference.
  • Owned data curation end to end — self-hosted large annotation models on GCP and Vultr (e.g., Qwen3.5-122B-A10B-FP8, DeepSeek-R1) with OmniParser to build chain-of-thought datasets of multiple types (next-step prediction, intent/goal-based, QA for GRPO).
  • Trained with DDP, then switched to tensor-parallel training after finding longer context improved accuracy, and applied GRPO reinforcement learning (LLM-as-judge reward) to improve reasoning and accuracy.
  • Deployed the 32B model on GCP with vLLM tensor parallelism behind OpenAI-compatible endpoints, and built a FastAPI chat UI with streaming for stakeholder evaluation.
  • Built an execution mesh on the OpenKoda platform to orchestrate agent-driven claim-filing workflows, hosting coder models to translate planned steps into concrete actions on the application.
  • Built Snowflake SQL pipelines to select MSK radiograph cohorts from DICOM metadata and CPT codes, identifying ~469K usable studies across multiple providers.
  • Ran cohort-level EDA and developed view-position normalization and image-selection logic to prepare datasets for VLM training and evaluation.

Associate Software Developer - Omegasoft Technologies Pvt. Ltd.

(2022-08 - 2023-07)

Built and maintained backend features in C# for the reports section of a US healthcare (palliative-care) product.

  • Wrote new LINQ queries for reports, and debugged and optimized existing queries to improve performance and reliability.

Education

M.Tech - Computer Science - IIT Kharagpur (2025)

B.E - Computer Science - Savitribai Phule Pune University (2022)

HSC (Class XII) - Maharashtra State Board (2017)

SSC (Class X) - Maharashtra State Board (2015)

Skills

Data Scientist

Reviews

Related