AI/ML Research Engineer

Akash Pawar

Transforming Complex AI into Elegant Solutions

AI engineer who bridges theory and application across domains — from fine-tuning mathematical reasoning models and vision models to protein bioinformatics and large-scale graph analysis. Specializes in rapid domain mastery and building systems that solve real problems with intelligent design.

Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
Python
PyTorch
SQL
HuggingFace
AWS
LangChain
Docker
FastAPI
Tableau
LoRA
MS
Machine Learning
3.93
GPA
4+
Years AI/ML Experience
Explore Project Nodes Initialize Contact
NEURAL_PROFILE.config

About Me

Network Configuration Overview

I am an AI/ML Research Engineer with expertise in Deep Learning, NLP, and traditional Machine Learning, driven by a passion for solving complex problems with intelligent systems. I have completed my Master's in Machine Learning from Stevens Institute of Technology, with a 3.93 GPA, where I honed my skills in both theoretical foundations and real-world applications.

I have built and optimized machine learning systems across NLP, deep learning, and computational biology, tackling challenges in text classification, retrieval-augmented generation, large-scale model fine-tuning, and predictive analytics. I thrive in environments where rapid learning and adaptation are key.

With a strong foundation in research and hands-on implementation, I am always eager to tackle new challenges, quickly learn emerging technologies, and refine complex models into production-ready solutions. My ability to adapt to new domains while optimizing performance makes me a valuable asset in any machine learning-driven environment.

Name Akash Pawar
Email akashpawar9619@gmail.com
Location Jersey City, NJ, USA
Degree MS in Machine Learning
Akash Pawar
Machine Learning CV
EXPERIENCE_LOG.dat

Work Experience

Model Training & Optimization Journey

Artificial Intelligence and Machine Learning Research Assistant

Stevens Institute of Technology
April 2025 - Present
Hoboken, New Jersey
95%+ reduction in grading time
GPT-based evaluation system
  • Designed and implemented an agentic LangGraph-based evaluation system using GPT models, achieving 95% reduction in grading time while maintaining assessment quality.
  • Engineered modular framework with conditional routing for comprehensive code assessment across syntax, requirements fulfillment, analysis and visualization quality.
  • Developed robust state management architecture with Pydantic schemas for structured LLM outputs, enabling consistent JSON feedback generation for learning management system integration.
  • Built automated evaluation pipelines that process student submissions through multi-stage assessment workflows, significantly reducing manual grading overhead for academic courses.

Data Analytics & Machine Learning Fellow Trainee

ElevateMe
March 2025 - Present
Remote
Increased R² from 43.8% → 80.0%
Average MAE of $170
Median Error of $39
Entertainment Revenue Prediction Model
  • Engineered a machine learning pipeline achieving 80.0% R² score for event gross revenue prediction, progressing systematically through 6 development phases with comprehensive feature engineering that transformed baseline performance from 43.8% to production level predictive capability.
  • Built production-ready data pipelines processing 305K+ entertainment contracts across multiple relational tables, implementing systematic feature engineering including artist experience tiers, genre pricing intelligence, financial commission analysis, and artist pricing anchors that revealed critical revenue drivers.
  • Discovered breakthrough pricing intelligence patterns through multi-table joins and financial feature engineering, identifying that multi-artist events average 2.3x pricing multipliers and commission data serves as the strongest revenue predictor, enabling precise contract pricing within $170 accuracy.
  • Developed comprehensive ML evaluation framework with cross-validation pipelines and automated feature selection, optimizing XGBoost algorithms through hyperparameter tuning and implementing smart feature pruning that reduced overfitting while maintaining 80.0% prediction accuracy for business-critical pricing decisions.

Master's Student - Advanced Projects

Stevens Institute of Technology
Sep 2023 - Dec 2024
Hoboken, New Jersey
Critical Node Identification
RL Proof of concept
Time series Forecasting
Topic Modeling and Sentiment Research
Sentiment Analysis: VADER vs BERT

Dependency Chain Analysis

Neo4j • NetworkX • ML • RL
  • Led a team of five in developing a system for Feature-Driven Critical Node Identification and Dependency Chain Optimization using Reinforcement Learning.
  • Designed a hybrid scoring function integrating topological (degree centrality, betweenness, PageRank) and semantic (node types, dependency scopes, quality indicators) features for critical node identification.
  • Designed an adaptive data sampling approach, efficiently extracting a representative dataset from a 15M-node and 178M-edge dependency graph.
  • Developed graph processing pipelines using Neo4j, NetworkX, and Python, allowing scalable data ingestion, feature extraction, and embedding generation.
  • Processed a large-scale dependency graph with 442,275 nodes and 499,760 edges, implementing Node2Vec embeddings for structural learning.
  • Achieved 100% precision and recall in critical node classification using a Random Forest classifier, significantly outperforming baseline methods.
  • Built a reinforcement learning environment with a custom reward function to optimize dependency chains, ensuring security, performance, and freshness.
  • Implemented PPO (Proximal Policy Optimization) for reinforcement learning, leveraging a multi-head attention model for dynamic chain optimization.
  • Evaluated system performance on dependency characteristics, classification accuracy, and RL-based chain generation, demonstrating potential for real-world dependency management.
  • Conducted extensive experiments on dependency chain properties, RL training strategies, and scoring methodologies, ensuring scalability and robustness.

Temporal Analysis of COVID-19 Sentiments

PyTorch • BERTopic • Prophet
  • Objective: Analyzed public discourse and emotional responses to COVID-19 through 400,000+ tweets across three phases of the pandemic.
  • Sentiment & Emotion Analysis: Combined traditional sentiment analysis with a fine-tuned BERT model trained on 10 emotion categories, including optimism, anxiety, and denial.
  • Topic Modeling: Employed BERTopic and HDBSCAN clustering to track the evolution of public concerns, from crisis response to vaccine-related discussions.
  • Geographical Sentiment Trends: Identified regional sentiment variations, with U.S. tweets showing the highest negativity and Indian tweets reflecting a slightly positive bias.
  • Forecasting Public Sentiment: Used Prophet time-series forecasting to analyze sentiment evolution and predict future trends in public discourse.
  • Data Processing & Preprocessing: Cleaned and structured 411,887 tweets, implementing custom regex-based text cleaning, filtering, and missing-value handling.
  • Temporal Insights: Revealed shifts from crisis-focused discussions to political debates and vaccine discourse, showcasing the dynamic evolution of public sentiment.
  • Multi-Dimensional NLP Pipeline: Integrated deep learning, topic modeling, and time-series forecasting for a comprehensive understanding of societal reactions to global crises.
  • Crisis Communication Implications: Provided actionable insights into how governments and public health officials can tailor communication strategies for effective crisis management.

Bachelor's Student - Capstone Project

Mumbai University
Aug 2019 - May 2023
Mumbai, India
98.5% accuracy detecting depression
Sentiment detection and classification
98.2% accuracy classifying type of sentiment
Context aware system

Depression Amelioration Chatbot

RoBERTa • SLM • Chatbot • Tensforflow
  • Led a team of four in developing a deep learning-based conversational agent for mental health intervention.
  • Designed a two-stage RoBERTa-based classification system for depression detection with 98.5% accuracy.
  • Implemented a multi-class classification model for mental health categorization, achieving 98.2% accuracy.
  • Integrated BlenderBot-400M for context-aware response generation in general conversations.
  • Developed a curated response database covering 80 unique intents to enhance chatbot reliability.
  • Engineered a conversational context management system to maintain coherence across interactions.
  • Deployed the chatbot on Discord for accessibility, anonymity, and real-time mental health support.
  • Conducted a peer-based review demonstrating a significant reduction in depressive symptoms among users.
  • Ensured ethical considerations by prioritizing data privacy and non-judgmental, empathetic responses.
  • Highlighted the potential of AI-driven mental health interventions for scalable and accessible support.
PROJECT_NETWORKS.json

Featured Projects

Neural Architectures in Action

DeepSeek Mathematical Reasoning

DeepSeek Mathematical Reasoning

Fine-tuned model outperforming Claude-3.5 Sonnet with 98.8% parameter reduction

Repository

DeepSeek Mathematical Reasoning

⚡ Outperformed Claude-3.5 📊 98.8% Parameter Reduction 🚀 2x Inference Speed

Fine-tuned DeepSeek-R1-Distill-Qwen-1.5B model that outperformed Claude-3.5 Sonnet on mathematical reasoning tasks. Reduced trainable parameters by 98.8% while doubling inference speed through LoRA adaptation and Unsloth framework optimization.

LoRA Unsloth LLM Fine-tuning Mathematical Reasoning
Protein Subcellular Localization

Protein Localization Predictor

High-accuracy protein localization using Meta AI's ESM2-3B with 92.09% top-5 accuracy

Repository

Protein Subcellular Localization Predictor

🎯 92.09% Top-5 Accuracy ⚡ 1.8x Training Speed 💾 50% Memory Reduction

High-accuracy protein localization system using Meta AI's ESM2-3B, achieving 84.79% top-3 and 92.09% top-5 accuracy across 12 cellular locations. Optimized training performance through mixed precision and gradient checkpointing.

ESM2 PyTorch Bioinformatics Mixed Precision
Advanced LLM Fine-tuning

LLM Data Curation Pipeline

Production-ready LLM optimization using NVIDIA's intelligent data curation achieving significant performance gains

Repository

Advanced LLM Fine-tuning with Data Curation

📉 6.7% Loss Reduction 🔥 NVIDIA NeMo Curator ⚡ Resource Optimized

Built production-ready LLM optimization pipeline leveraging NVIDIA NeMo Curator's advanced data filtering algorithms. Implemented intelligent prompt complexity classification and task-aware data selection, achieving 6.7% validation loss reduction while maintaining training efficiency on resource-constrained hardware through sophisticated curation strategies.

NVIDIA NeMo Curator QLoRA Data Quality Filtering Prompt Classification
RAG Document QA System

RAG Document QA System

Full-stack PDF document QA system with LangChain, ChromaDB, and React frontend

Repository

RAG Document QA System

🔍 Vector Search with ChromaDB 📄 PDF Processing Pipeline 🎯 95% Query Accuracy

Built production-ready document QA system enabling natural language queries on PDF documents. Features secure file upload, intelligent text chunking, vector embeddings storage, and real-time retrieval-augmented generation with modern React interface.

LangChain React Flask ChromaDB OpenAI Tailwind CSS
Dependency Chain Analysis

Dependency Chain Analysis

Large-scale graph analysis with Neo4j achieving 100% F1-score in critical node identification

Repository

Dependency Chain Analysis

🎯 100% F1-Score ⚡ 25s Processing Time 📊 500K+ Relationships

Processed 500,000+ relationships across 442,275 nodes in large-scale dependency graph using Neo4j and NetworkX. Implemented Node2Vec embeddings combined with handcrafted features, achieving perfect classification metrics.

Neo4j NetworkX Node2Vec Random Forest
Interactive Neural Network Visualizer

Neural Network Visualizer

Real-time neural network visualization with live activations and SVG animations

Repository

Interactive Neural Network Visualizer

🖼️ 28x28 Grid ⚡ Real-time Viz 🎨 SVG Animation

Real-time neural network visualization system with React frontend and FastAPI backend. Features live visualization of hidden layer activations, inter-layer connections via SVG, and prediction probabilities as users draw digits on interactive grid.

React FastAPI Real-time Viz SVG Animation
StackOverflow Questions Ranking System

StackOverflow Questions Ranking

Multi-label classification and intelligent ranking system using BERT transformers with NDCG evaluation

Repository

StackOverflow Questions Ranking System

🎯 NDCG Evaluation 🏷️ 50 Programming Tags 🤖 BERT Classification

Built sophisticated multi-label classification system using BERT transformers to automatically categorize and rank StackOverflow questions across 50 programming language tags. Implemented NDCG (Normalized Discounted Cumulative Gain) evaluation with BCEWithLogitsLoss for optimal ranking performance, enabling intelligent question prioritization and tag-based relevance scoring.

BERT Multi-label Classification NDCG Ranking PyTorch
Neural Machine Translation System

Neural Machine Translation

English-to-Spanish translation using bidirectional LSTM encoder-decoder architecture

Repository

Neural Machine Translation

🎯 Teacher Forcing 🔄 Bidirectional LSTM 📊 BLEU Evaluation

English-to-Spanish neural machine translation using bidirectional LSTM encoder-decoder architecture. Implements character-level tokenization with temperature-based multinomial sampling for improved translation diversity and quality assessment via BLEU metrics.

Bidirectional LSTM Seq2Seq BLEU Evaluation Teacher Forcing
Machine Learning from Scratch

ML from Scratch

Complete 2-layer neural network with manual backpropagation using only NumPy

Repository

Machine Learning from Scratch

⚡ Xavier Init 🧠 Manual Backprop 🔢 NumPy Only

Implemented complete 2-layer neural network with manual backpropagation using only NumPy, tested on MNIST and Fashion-MNIST datasets. Features Xavier initialization, numerically stable softmax, and custom learning rate scheduling.

NumPy Manual Backprop MNIST Xavier Init
Explore More Projects
SKILL_MATRIX.cfg

Technical Skills

Feature Space & Computational Stack

LLM & Neural Networks

PyTorch
Hugging Face
LangChain
LangGraph
LangSmith
LoRA/QLoRA
ESM2
Unsloth
TRL
PEFT
BitsAndBytes
Attention Mechanisms
BERT
GPT
Transformers
RLHF

MLOps & Cloud

Docker
AWS
Azure ML
FastAPI
Flask
Git
MLflow
Weights & Biases
Gradio
AWS S3
AWS Lambda
SageMaker
Mixed Precision
Gradient Checkpointing
KV-Cache

ML & Deep Learning

TensorFlow
Scikit-learn
XGBoost
LightGBM
Random Forest
FLAML
CNN
RNN/LSTM
GRU
Autoencoders
Reinforcement Learning
Transfer Learning

Databases & Data Engineering

MySQL
Feature Engineering
Neo4j
DuckDB
Chroma
Pinecone
Vector Databases
ETL Pipelines
Data Preprocessing

Programming & Frameworks

Python
JavaScript
NumPy
Pandas
Dask
Pydantic
Regex
JSON
Subprocess
React
TailwindCSS
Vite

Visualization & Analytics

Seaborn
Matplotlib
Plotly
Streamlit
Tableau
NetworkX
SVG Animations
Real-time Viz
Interactive Dashboards

NLP & Text Processing

OpenAI
spaCy
NLTK
BERTopic
TF-IDF
BM25
Word2Vec
Sentence Transformers
Text Mining
Sentiment Analysis
Tokenization
Prophet

Specialized Techniques

Computer Vision
Image Segmentation
Time Series Analysis
Graph Analysis
Node2Vec
Protein Bioinformatics
Statistical ML
Hyperparameter Optimization
Cross Validation
A/B Testing
Ensemble Methods
SMOTE
CONNECTION_LAYER.init

Initialize Contact

Establish Neural Pathways

Neural Mail

akashpawar9619@gmail.com

Connect

Professional Network

LinkedIn Profile

Connect

Code Repository

GitHub Projects

Explore

Ready to push the boundaries of AI together?