HARSH VARDHAN VERMA profile photo 1 on Tal

HARSH VARDHAN VERMA

MLE-II @ Pibit.ai | IIT Roorkee’23

Bengaluru

3+ years of work experience

Pibit.ai
Aug 20251 yr 1 moPresent

Bengaluru, Karnataka, India

AI-Powered Loss Run Processing • Led the development of Loss Run GPT, an LLM-powered document processing pipeline from the ground up, replacing an unsustainable legacy pipeline built on ad-hoc, template-specific code fixes, and proving the viability of the GPT-based approach over CV/rule-based heuristics. • Led the optimization of Loss Run GPT for Straight-Through Processing (STP), revamping the Template Mapper and engineering automated pre-processing for edge cases (blurred/dense scans, multi-page claims) — scaling coverage to 800+ templates across 25+ production clients while processing 6,000–8,000 documents/month at 90%+ extraction accuracy. • Drove a 50–60% improvement in Graph Edit Distance and a 94% field match ratio through systematic testing-to-deployment optimization, cutting new-template onboarding time from 2 weeks to 2 hours and enabling fully automated STP processing. • Proposed and built the Loss Run Prompt Registry, a centralized, database-driven platform enabling field-level prompt versioning and testing; expanded its scope into an operations hub covering Template Mapper configs, quality-check tracking, and task monitoring — driving a 3x jump in template throughput (5–10 to 20–30 templates/day). • Built an automated quality assurance system combining validation rules with a separate post-extraction pipeline for deduplication and business logics based on direct collaboration with Operations and QA teams.
KeepFlying®
Jul 20232 yrs 1 moJul 2025

Led end-to-end inventory optimization for an airline client using Python, PySpark, SQL, and Databricks, integrating diverse datasets. Built automated ETL pipelines with LakeFlow and Spark to process and refresh data daily. Developed models for safety stock, EOQ, reorder points, FSN-ABC-XYZ classification, and vendor rating to drive data-informed procurement. Designed and implemented an RAG system to manage and query large-scale airworthiness documents across multiple aircraft and document types. Built an automated document processing pipeline using OCR from Azure AI Document Intelligence, LangChain, and BAAI/bge-small-en-v1.5 embeddings for document parsing, chunking, and semantic indexing via Azure AI Search. Engineered a custom hybrid retrieval mechanism combining semantic and keyword search for improved query accuracy and coverage. Integrated GPT-4o for answer generation with image-based context input and user feedback mechanisms. Developed hybrid document type identification modules using regex patterns and LLM-based few-shot and role prompting (LLaMA 3.1, GPT-3.5, GPT-4o), achieving 95–98% classification accuracy across text and image inputs for various document types. Constructed document information extraction pipelines using Azure AI Document Intelligence and prompt-engineered LLMs, achieving 90- 95% structured extraction accuracy across text and image inputs for various document types. Built an evaluation workflow in MLflow, including a custom metric for assessing structured JSON extraction accuracy, enabling automated benchmarking across models, prompts and document types.
KeepFlying®
Jan 20236 mosJun 2023

KeepFlying®
May 20223 mosJul 2022

The Sparks Foundation
Nov 20211 moNov 2021

Technocolabs Softwares
Oct 20212 mosNov 2021

GitHub
@hv129

HARSH’s AI Stack

antigravity
antigravity
codex
codex
opencode
opencode

Education

Indian Institute of Technology, Roorkee

Integrated M.Tech., Geophysical Technology

Indian Institute of Technology, Roorkee

20185 yrs2023
Indian Institute of Technology, Roorkee

Master's Degree

Indian Institute of Technology, Roorkee

Join HARSH and 10k+ other professionals on

Tal Boss

Tal is the app where bosses meet engineers & builders directly, without middlemen

Download Tal Boss

HARSH VARDHAN VERMA on Tal

Headline

MLE-II @ Pibit.ai | IIT Roorkee’23

About

Graduated from IIT Roorkee, currently contributing as a Data Scientist with 2+ years of experience. Worked alongside teams to develop AI-driven solutions leveraging advanced analytics, machine learning, Generative AI, and optimization techniques. Proven ability to drive tangible impact through data-driven solutions in Inventory Optimization, Retrieval Augmented Generation (RAG) systems, Generative AI, Prompt Engineering, Statistical Modeling, Resource Optimization and Structured Data Extraction. Possesses strong technical skills in Python, SQL, machine learning libraries, LLMs, and cloud technologies.

Location

Bengaluru

AI stack

antigravity

codex

opencode

Work experience

Machine Learning Engineer at Pibit.ai, Aug 2025 - Present · 1 yr 1 mo, Bengaluru, Karnataka, India, AI-Powered Loss Run Processing • Led the development of Loss Run GPT, an LLM-powered document processing pipeline from the ground up, replacing an unsustainable legacy pipeline built on ad-hoc, template-specific code fixes, and proving the viability of the GPT-based approach over CV/rule-based heuristics. • Led the optimization of Loss Run GPT for Straight-Through Processing (STP), revamping the Template Mapper and engineering automated pre-processing for edge cases (blurred/dense scans, multi-page claims) — scaling coverage to 800+ templates across 25+ production clients while processing 6,000–8,000 documents/month at 90%+ extraction accuracy. • Drove a 50–60% improvement in Graph Edit Distance and a 94% field match ratio through systematic testing-to-deployment optimization, cutting new-template onboarding time from 2 weeks to 2 hours and enabling fully automated STP processing. • Proposed and built the Loss Run Prompt Registry, a centralized, database-driven platform enabling field-level prompt versioning and testing; expanded its scope into an operations hub covering Template Mapper configs, quality-check tracking, and task monitoring — driving a 3x jump in template throughput (5–10 to 20–30 templates/day). • Built an automated quality assurance system combining validation rules with a separate post-extraction pipeline for deduplication and business logics based on direct collaboration with Operations and QA teams.

Data Scientist at KeepFlying®, Jul 2023 - Jul 2025 · 2 yrs 1 mo, Led end-to-end inventory optimization for an airline client using Python, PySpark, SQL, and Databricks, integrating diverse datasets. Built automated ETL pipelines with LakeFlow and Spark to process and refresh data daily. Developed models for safety stock, EOQ, reorder points, FSN-ABC-XYZ classification, and vendor rating to drive data-informed procurement. Designed and implemented an RAG system to manage and query large-scale airworthiness documents across multiple aircraft and document types. Built an automated document processing pipeline using OCR from Azure AI Document Intelligence, LangChain, and BAAI/bge-small-en-v1.5 embeddings for document parsing, chunking, and semantic indexing via Azure AI Search. Engineered a custom hybrid retrieval mechanism combining semantic and keyword search for improved query accuracy and coverage. Integrated GPT-4o for answer generation with image-based context input and user feedback mechanisms. Developed hybrid document type identification modules using regex patterns and LLM-based few-shot and role prompting (LLaMA 3.1, GPT-3.5, GPT-4o), achieving 95–98% classification accuracy across text and image inputs for various document types. Constructed document information extraction pipelines using Azure AI Document Intelligence and prompt-engineered LLMs, achieving 90- 95% structured extraction accuracy across text and image inputs for various document types. Built an evaluation workflow in MLflow, including a custom metric for assessing structured JSON extraction accuracy, enabling automated benchmarking across models, prompts and document types.

Data Science Intern at KeepFlying®, Jan 2023 - Jun 2023 · 6 mos

Data Science Intern at KeepFlying®, May 2022 - Jul 2022 · 3 mos

Data Science and Business Analytics at The Sparks Foundation, Nov 2021 - Nov 2021 · 1 mo

Data Science Intern at Technocolabs Softwares, Oct 2021 - Nov 2021 · 2 mos

Education

Integrated M.Tech., Geophysical Technology, Indian Institute of Technology, Roorkee

Master's Degree, Indian Institute of Technology, Roorkee

Public links

github: https://github.com/hv129