Aditya Aggrawal profile photo 1 on Tal

Aditya Aggrawal

Data Scientist II at Eli Lilly and Company

Greater Bengaluru Area

5+ years of work experience

Eli Lilly and Company
Dec 20241 yr 10 mosPresent

LEDR — Local Exhibit Document Review • Built an AI-powered document-review tool for an internal compliance team, cutting review time from 2.5 hours to under 10 minutes per submission by automating extraction and policy evaluation across cost centres, associate details, and approval chains. • Designed a two-stage pipeline: AWS Textract for OCR and layout-aware extraction feeding an LLM layer for structured parsing and policy-rule evaluation. • Fine-tuned Llama 70B with QLoRA on domain-specific compliance data to replace GPT-4o at inference keeping data in-house and reducing per-inference cost at scale while benchmarking to comparable accuracy vs. GPT-4o on a labeled test set.
Cognizant
Nov 20222 yrs 2 mosDec 2024

Personalised Learning Recommender Chatbot Multi-Agent Platform • Architected a supervisor + specialist multi-agent and agentic RAG system using pgvector on a LangChain ReAct framework, a routing agent decomposes compound, cross-domain queries, dispatches them to specialist sub-agents, and synthesizes a unified response handling multi-domain questions in a single turn. • Served the platform via FastAPI, containerized with Docker, with Azure Redis persisting conversation context across turns sustaining 5,100+ daily queries at 82% CSAT under a p90 latency < 5s. • Led a team of 5 engineers to design and ship the platform end-to-end. • Built a clarification loop that elicits missing context (role, skill level, goals) before generating recommendations, improving relevance and reducing back-and-forth. • Deflected 9,000 deflected support tickets and 16,000+ person-hours saved annually.
Cognizant
Jun 20221 yr 7 mosDec 2023

• Benchmarked Word2Vec, TF-IDF, BERT, and Longformer for semantic similarity on legal case documents; selected Longformer for its best performance on held-out expert-annotated pairs. • Built a document-embedding pipeline for legal cases exceeding 12,000 tokens (beyond BERT’s 512-token limit) using hierarchical chunking; achieved 0.75 Pearson correlation with legal-expert similarity scores. • Trained a 17-class NER model on legal corpora; Legal BERT reached 91% macro-F1 vs. BiLSTM-CRF baselines.
Cognizant
Nov 20211 yr 1 moNov 2022

Canspirit Artificial Intelligence
Mar 20205 mosJul 2020

Pune, Maharashtra, India

1. Implemented vehicle detection and localization in aerial images using semantic segmentation using U-Net a state-of-the-art computer vision algorithm based on CNN, obtained a remarkable F1-score of 0.82. 2. Improved and optimized other computer vision algorithms using different techniques like hyperparameter tuning, k-fold cross-validation, and data augmentation.
Cuemath Pune
Apr 20182 mosMay 2018

1. Performed the responsibility of a sales representative on behalf of Cuemath to convert dead teacher leads into active. 2. Successfully converted a total of 15 homemakers into a Cuemath teacher partner.

Aditya’s AI Stack

hugging face
hugging face
langchain
langchain
langgraph
langgraph
meta
meta
openai
openai
Chroma
Chroma

Education

Maharashtra Institute of Technology

Bachelor of Technology

Maharashtra Institute of Technology

20174 yrs2021

Join Aditya and 10k+ other professionals on

Tal Boss

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

Download Tal Boss

Aditya Aggrawal on Tal

Headline

Data Scientist II at Eli Lilly and Company

Location

Greater Bengaluru Area

AI stack

hugging face

langchain

langgraph

meta

openai

Chroma

Work experience

Data Scientist II at Eli Lilly and Company, Dec 2024 - Present · 1 yr 10 mos, LEDR — Local Exhibit Document Review • Built an AI-powered document-review tool for an internal compliance team, cutting review time from 2.5 hours to under 10 minutes per submission by automating extraction and policy evaluation across cost centres, associate details, and approval chains. • Designed a two-stage pipeline: AWS Textract for OCR and layout-aware extraction feeding an LLM layer for structured parsing and policy-rule evaluation. • Fine-tuned Llama 70B with QLoRA on domain-specific compliance data to replace GPT-4o at inference keeping data in-house and reducing per-inference cost at scale while benchmarking to comparable accuracy vs. GPT-4o on a labeled test set.

AI/ML Engineer at Cognizant, Nov 2022 - Dec 2024 · 2 yrs 2 mos, Personalised Learning Recommender Chatbot Multi-Agent Platform • Architected a supervisor + specialist multi-agent and agentic RAG system using pgvector on a LangChain ReAct framework, a routing agent decomposes compound, cross-domain queries, dispatches them to specialist sub-agents, and synthesizes a unified response handling multi-domain questions in a single turn. • Served the platform via FastAPI, containerized with Docker, with Azure Redis persisting conversation context across turns sustaining 5,100+ daily queries at 82% CSAT under a p90 latency < 5s. • Led a team of 5 engineers to design and ship the platform end-to-end. • Built a clarification loop that elicits missing context (role, skill level, goals) before generating recommendations, improving relevance and reducing back-and-forth. • Deflected 9,000 deflected support tickets and 16,000+ person-hours saved annually.

Junior AI Engineer at Cognizant, Jun 2022 - Dec 2023 · 1 yr 7 mos, • Benchmarked Word2Vec, TF-IDF, BERT, and Longformer for semantic similarity on legal case documents; selected Longformer for its best performance on held-out expert-annotated pairs. • Built a document-embedding pipeline for legal cases exceeding 12,000 tokens (beyond BERT’s 512-token limit) using hierarchical chunking; achieved 0.75 Pearson correlation with legal-expert similarity scores. • Trained a 17-class NER model on legal corpora; Legal BERT reached 91% macro-F1 vs. BiLSTM-CRF baselines.

Programmer Analyst Trainee at Cognizant, Nov 2021 - Nov 2022 · 1 yr 1 mo

Machine learning at Canspirit Artificial Intelligence, Mar 2020 - Jul 2020 · 5 mos, Pune, Maharashtra, India, 1. Implemented vehicle detection and localization in aerial images using semantic segmentation using U-Net a state-of-the-art computer vision algorithm based on CNN, obtained a remarkable F1-score of 0.82. 2. Improved and optimized other computer vision algorithms using different techniques like hyperparameter tuning, k-fold cross-validation, and data augmentation.

Sales And Marketing at Cuemath Pune, Apr 2018 - May 2018 · 2 mos, 1. Performed the responsibility of a sales representative on behalf of Cuemath to convert dead teacher leads into active. 2. Successfully converted a total of 15 homemakers into a Cuemath teacher partner.

Education

Bachelor of Technology, Maharashtra Institute of Technology