Mahesh Deshwal profile photo 1 on Tal

Mahesh Deshwal

Senior AI Research Engineer || NLP (Gen AI, LLM Agents as you might know) | Deep Learning | Currently working on AI Evals, Alignment, Safety, Trust & Responsibilities: Basically guardrails & all the hard part remaining

India

7+ years of work experience

Domyn
Nov 20241 yr 6 mosPresent

Remote

Building AI Sentinel: Working on AI Evaluation & Security Research (Hallucination is my fav) And then building Foundational Models for regulated industries (Finance is the current one for now) Improving in house Multi Agentic System (Crystal to be precise)
Chegg Inc.
Sep 20231 yr 3 mosNov 2024

New Delhi, Delhi, India

Designed and deployed an LLM evaluation framework to detect hallucinations and Evaluate responses against custom rules in real-time using hybrid approach of rules, LLM-as-Judge, Fine tuned LLMs and DeBERTa classifiers. Enhanced the platform by integrating subject-specific models through fine-tuning LLMs and OpenAI models with Retrieval-Augmented Generation (RAG). Boosted search and recommendation result quality by fine-tuning a re-ranker model on top of GTE-large.
Jackett
Mar 20221 yr 7 mosSep 2023

Singapore

Owned an end-to-end document digitization OCR service in multipe countries using Faster RCNN with Transformer as backbone increasing internal content creation efficiency by amost double Deployed an OCR pre-processing pipeline for document image column detection, header-footer removal, skew correction, noise reduction to reduce QC effort Developed a robust pipeline for automated document image classification using a fine-tuned quantized DiT, heuristics, metadata, and traditional CV approaches
Vedantu
Oct 20216 mosMar 2022

Bangalore Urban, Karnataka, India

Successfully submitted POCs for replacing third part APIs dependencies to build 2 different in house OCRs. Generated SMILES string from organic compounds diagram and LaTex equations from Maths using CNN - Transformer based Encode-Decoder architecture. Improved overall search and diagram search using a two stage pipeline to generate crops of diagrams in images using Scaled YOLOv4 followed by MobileNet fine tuned as Siamese with Triplet Loss to generate embedding for image similarity.
Vedantu
Jan 20192 yrs 9 mosSep 2021

Bengaluru, Karnataka, India

After InstaSolv merged into Vedantu, I worked on: Envisioned and successfully deployed a Text-Video and Text-Text recommendation engine to increase user engagement and retention after fine tuning a DistilBERT model on extended vocabulary for PCMB domain using approach from paper exBERT. Cleaned the entire database of 18M+ documents for grouping of fragmented instances and removing unwanted ones from the group using text similarity hashing (Sim Hashing) along with image embedding. Built a XGBoost model to predict chapter of a text question from a pool of 250 NCERT chapters on highly imbalanced data with F-1 score of 86%.
GitHub
@deshwalmahesh

Education

Master's degree, Computer Engineering

YMCA University of Science & Technology

20163 yrs2019
Maharaja Surajmal Institute Of Technology

Bachelor's degree, Information Technology

Maharaja Surajmal Institute Of Technology

20124 yrs2016

Join Mahesh and 10k+ other professionals on

Tal Boss

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

Download Tal Boss

Mahesh Deshwal on Tal

Headline

Senior AI Research Engineer || NLP (Gen AI, LLM Agents as you might know) | Deep Learning | Currently working on AI Evals, Alignment, Safety, Trust & Responsibilities: Basically guardrails & all the hard part remaining

About

Data Scientist, AI researcher (author of paper PHUDGE) and Open Source Contributor with 6.5+ years of experience in solving end user business problems using ML and AI extensively in the field of Computer Vision, NLP, Gen AI, Recommendation and personalisation. Quite experienced in serving end to end solutions starting from user requirements to deploying the scalable models using MLOPS. Currently working with Agents. Most of my life I've worked with 0-1 stage companies and so it is evident from my working experience, I'm more of a startup guy who loves to deliver business values using AI and I've thrived in roles where I could work on a whole lot of things end to end from profiling users, data collection, cleaning, R&D, model development and most importantly deployment and monitoring like the intense growth roles in 0 -1 phase startups. My Paper: https://github.com/deshwalmahesh/PHUDGE Major open source collaborations include Google Research: MAXIM and Huawei Noah's Ark Lab: CURL. Second research paper available at: http://www.jetir.org/papers/JETIR1905E18.pdf Recent AI blog at: https://medium.com/p/13f842bb77ce

Location

India

Work experience

Senior AI Research Engineer at Domyn, Nov 2024 - Present · 1 yr 6 mos, Remote, Building AI Sentinel: Working on AI Evaluation & Security Research (Hallucination is my fav) And then building Foundational Models for regulated industries (Finance is the current one for now) Improving in house Multi Agentic System (Crystal to be precise)

Senior ML Engineer at Chegg Inc., Sep 2023 - Nov 2024 · 1 yr 3 mos, New Delhi, Delhi, India, Designed and deployed an LLM evaluation framework to detect hallucinations and Evaluate responses against custom rules in real-time using hybrid approach of rules, LLM-as-Judge, Fine tuned LLMs and DeBERTa classifiers. Enhanced the platform by integrating subject-specific models through fine-tuning LLMs and OpenAI models with Retrieval-Augmented Generation (RAG). Boosted search and recommendation result quality by fine-tuning a re-ranker model on top of GTE-large.

Founding Senior Data Scientist at Jackett, Mar 2022 - Sep 2023 · 1 yr 7 mos, Singapore, Owned an end-to-end document digitization OCR service in multipe countries using Faster RCNN with Transformer as backbone increasing internal content creation efficiency by amost double Deployed an OCR pre-processing pipeline for document image column detection, header-footer removal, skew correction, noise reduction to reduce QC effort Developed a robust pipeline for automated document image classification using a fine-tuned quantized DiT, heuristics, metadata, and traditional CV approaches

Software Engineer II at Vedantu, Oct 2021 - Mar 2022 · 6 mos, Bangalore Urban, Karnataka, India, Successfully submitted POCs for replacing third part APIs dependencies to build 2 different in house OCRs. Generated SMILES string from organic compounds diagram and LaTex equations from Maths using CNN - Transformer based Encode-Decoder architecture. Improved overall search and diagram search using a two stage pipeline to generate crops of diagrams in images using Scaled YOLOv4 followed by MobileNet fine tuned as Siamese with Triplet Loss to generate embedding for image similarity.

Founding Machine Learning Researcher at Vedantu, Jan 2019 - Sep 2021 · 2 yrs 9 mos, Bengaluru, Karnataka, India, After InstaSolv merged into Vedantu, I worked on: Envisioned and successfully deployed a Text-Video and Text-Text recommendation engine to increase user engagement and retention after fine tuning a DistilBERT model on extended vocabulary for PCMB domain using approach from paper exBERT. Cleaned the entire database of 18M+ documents for grouping of fragmented instances and removing unwanted ones from the group using text similarity hashing (Sim Hashing) along with image embedding. Built a XGBoost model to predict chapter of a text question from a pool of 250 NCERT chapters on highly imbalanced data with F-1 score of 86%.

Education

Master's degree, Computer Engineering, YMCA University of Science & Technology

Bachelor's degree, Information Technology, Maharaja Surajmal Institute Of Technology

Public links

github: https://github.com/deshwalmahesh

Mahesh Deshwal - Senior AI Research Engineer in India | Tal