Fusing speech and text with Wav2Vec 2.0 and RoBERTa
Paresh Subedi, Data scientist & AI/ML engineer. I turn noisy data into decisions.
Open to Data & Analytics, ML Engineering and AI Systems roles.
(01)About
I make language models leaner, and data useful.
Graduate Research Assistant at Fordham University, researching neuroscience-inspired ways to make Transformer-based LLMs faster and more energy-efficient through inhibitory mechanisms and sparse activation. Before that, two years at Genpact in Bangalore managing SAP ERP accounts and building the Tableau reporting behind them.
- Model
- Paresh-v2026
- Architecture
- Data scientist + systems thinker
- Training data
- M.S. Data Science (Fordham), research in efficient Transformers, two years of ERP and analytics operations
- Intended use
- Analytics, ML engineering, AI product
- Known limitations
- will over-engineer a dashboard if left unsupervised
85%
Validation accuracy
CNN ensemble on HAM10000
1.3M
Rows engineered
A decade of daily S&P 500 data
10%
Top of cohort
Merit scholarship, Jain University
Education
- M.S. Data Science · Fordham University
- B.Tech (Honours), Computer Science, Data Science · Jain University
(02)Selected work
Work.
8 projects across research, deep learning, agents and analytics. Open one for the problem, the approach and a live diagram.
- 01data · 2026
- 02learning · 2025 — now
Neuroscience-inspired sparsity for Transformers
Research in progress
Open case → - 03output · 2023
A CNN ensemble on HAM10000, best final-year project
- 04learning · 2025
CrewAI agents that score and route engineering tickets
Shipped as a prototype
Open case → - 05output · 2025
A decade of prices, one interactive risk dashboard
1.3M
rows engineered from daily prices
Open case → - 06data · 2026
Deep learning on audio, with an agent that responds
In development
Open case → - 07output · 2026
Embeddings that cluster by moral foundation
Analysis in progress
Open case → - 08data · 2026
Grounded tutoring over a knowledge graph
Prototype
Open case →
(03)Skills
Toolkit.
Nine groups, ranked by how deep I have gone. Bars show relative strength, not years.
- 01
Python
- Python
- NumPy
- Pandas
- Matplotlib
- Seaborn
- 02
SQL
- SQL
- NoSQL
- Tableau
- Power BI
- 03
ML
- Scikit-learn
- XGBoost
- Feature engineering
- 04
Deep Learning
- PyTorch
- TensorFlow
- Keras
- CNN ensembles
- 05
NLP
- Hugging Face
- Embeddings
- Moral Integrity Corpus
- 06
LLMs & Agents
- CrewAI
- Multi-agent systems
- RAG
- Transformers
- 07
Cloud & MLOps
- Git
- Streamlit
- PySpark
- HDFS
- MapReduce
- 08
Data Viz
- Tableau
- Power BI
- Matplotlib
- Seaborn
- 09
Systems Architecture
- SAP
- Salesforce
- Pega
- Bash
(04)Experience
Path.
$ python train.py --career --epochs 2
01/02 · 2025 — Present
Fordham University
New York, USA
Graduate Research Assistant
- >Researching neuroscience-inspired methods to improve the energy efficiency and inference speed of Transformer-based LLMs, under faculty supervision.
- >Exploring inhibitory mechanisms and sparse activation to cut redundant computation in large-scale neural networks.
02/02 · 2023 — 2025
Genpact
Bangalore, India
Senior Associate
- >Onboarded and maintained customer accounts in SAP ERP systems, keeping data entry accurate and timely.
- >Resolved account discrepancies with cross-functional teams, reducing data errors and improving customer experience.
- >Built Tableau reports on customer account metrics that surfaced patterns improving operational efficiency.
- in production operations
- 2 yrs
checkpoints
M.S. Data Science
Fordham University · 2025 — Present
Graduate Research Assistant and Graduate Assistantship recipient.
B.Tech (Honours), Computer Science, Data Science
Jain University · 2019 — 2023
Coursework in Python, data analytics, machine learning, deep learning and NLP. Best Final-Year Project in deep learning for dermatology.