Bhagyesh
Data Analyst & Data Scientist
Data Analyst at Red Pico, based in Toronto. I make data trustworthy: validation, SQL, dashboards and applied ML.
I'm a data analyst with a background in applied machine learning. I care about the part most people skip: making sure the data is right before anyone builds on it.
At Red Pico I work on data quality: I build automated validation rules in Informatica, set up governance workflows in Dataiku, and tune SQL so downstream models and dashboards start from clean data.
Before that I worked on LLM training data at Innodata, and on NLP and machine learning projects at the Sheridan Centre for Applied AI. I graduated from Sheridan College with an Honours Bachelor of Computer Science in Data Analytics, and co-authored a paper on LLM-based topic modeling published at ACDSA 2025.
Where I've worked
March 2026 – Present
Data Analyst
Red Pico · Mississauga, ON
- Built automated validation rules in Informatica across 90% of critical pipelines in a single quarter, catching data errors early to support downstream ML modeling and business reporting.
- Built governance workflows in Dataiku to document model lineage, track bias metrics, and maintain compliance standards across deployed AI solutions.
- Optimized SQL queries and data transformations for tables with 100k+ rows, cutting runtimes and preparing clean datasets for internal dashboards.
August 2025 – March 2026
Generative AI Associate
Innodata Canada
- Improved LLM accuracy, fluency, and coherence by 25%+ by creating and validating training datasets using Python, NLP, and human-in-the-loop evaluation.
- Annotated and reviewed 10K+ text samples to align AI outputs with human values using data annotation workflows, prompt engineering, and Responsible AI practices.
- Reduced grammatical and structural errors by 30% through syntax analysis and NLP-based text quality checks AI-generated text.
- Accelerated model iteration cycles by 20% by collaborating with data scientists and engineers to refine LLM training specifications and data recipes, consistently meeting quality KPIs.
September 2024 – August 2025
AI/ML Developer
Sheridan Centre for Applied AI
- Developed an automated pipeline to extract fleet management KPIs from diverse plans using a locally deployed Meta Llama 3 LLM.
- Boosted KPI extraction accuracy and response relevance by 60% through advanced prompt engineering and few-shot learning, eliminating model retraining.
- Deployed BERTopic for automated topic modeling, revealing fleet management categories and streamlining taxonomy creation, improving efficiency by 70%.
- Built a custom fleet scorecard with NLP and ML, integrating an automated pipeline to generate data-driven KPI weights using Autoencoders.
May 2024 – August 2024
Machine Learning Data Scientist
Sheridan Centre for Applied AI
- Formulated optimized data cleaning scripts for analyzing 100,000+ patient records, leveraging parallel processing frameworks to handle large datasets efficiently, and extracting actionable insights to guide stakeholder presentations.
- Improved model accuracy by 40% by implementing a voting technique combining multiple machine learning models.
- Improved F1score and recall by 52.08% and 26.6% respectively in order to gain balanced performance matrix.
- Utilized synthetic data generation methods such as SMOTE and Random Oversampling to handle imbalanced datasets, doubling the recall and precision scores.
Things I've built
01LLM Eval Harness
Scores LLM answers for hallucination · leaderboard
→
- Built a harness that scores LLM answers for correctness, faithfulness, hallucination and relevance against a fixed question set with source passages, and ranks runs on a leaderboard.
- Tested two open models with two prompt styles on 40 SQuAD v2 questions, 12 of them with no answer in the passage: 160 answers in total.
- On the same 360M model, a grounded prompt cut hallucinations from 47.5% to 27.5%.
- The judge is a local NLI model plus embeddings, so it needs no API key. A hand check of 20 random answers agreed with the judge on 14 of 20.
September 2026
Python, NLI (DeBERTa v3), sentence embeddings, transformers.js, SQuAD v2
View on GitHub ↗
02DQ Rule Generate
Data quality rules from your data · live demo
→
- Built a tool that profiles a CSV (nulls, types, ranges, patterns) and suggests data quality rules from the data's own statistics, each with the reason it was suggested. No LLM or API key needed.
- Exports the same rule set as Great Expectations, pandera and SQL checks, so one review covers every tool in the pipeline.
- Validates new batches and flags failing rows with example values. On the sample data it caught all 10 planted problems.
- Runs fully in the browser with Pyodide: 33 rules from a 500 row sample in under a second, and about 2.5 seconds on 100,000 rows.
September 2026
Python, pandas, Pyodide, Great Expectations, pandera, SQL
Try the live demo ↗
View on GitHub ↗
03FairLens
AI hiring platform · 2nd best capstone
→
- Built FairLens, an AI-driven hiring platform using Python, Flask, and AWS to enable unbiased, DEI-aligned resume screening.
- Used LLMs for candidate-job matching and implemented PII anonymization to ensure fair, secure processing.
- Designed a real-time diversity dashboard in Python Dash to visualize workforce and recruitment funnel insights.
- Optimized performance and reduced costs by integrating OpenAI GPT-4o Mini and efficient AWS-based data handling.
September 2024 – April 2025
Python, Flask, AWS, Python Dash, OpenAI GPT-4o Mini
View on GitHub ↗
04AWS RAG Chatbot
90%+ relevance · under 500ms
→
- Built a document-aware chatbot using AWS and RAG architecture with Claude v2 LLM, achieving over 90% response relevance in domain-specific queries.
- Developed a scalable document pipeline using amazon.titan-embed-text-v1, generating 10,000+ vector embeddings from S3-stored files for high-speed retrieval from vector databases.
- Designed a web interface with AWS Amplify, supporting <500ms query response time and seamless user interaction across devices.
- Integrated OpenSearch for real-time similarity search, AWS Bedrock for LLM responses, and CloudWatch for monitoring, reducing debugging time by 40%.
August 2023
AWS Bedrock, Claude v2, Amazon Titan Embeddings, OpenSearch, S3, Amplify, CloudWatch
View on GitHub ↗
05LLM Topic Modeling
Code behind the ACDSA 2025 paper
→
- Code behind the ACDSA 2025 paper: improving BERTopic results on long documents with LLM-generated summaries.
06Fleet KPI Weight Agent
LangGraph KPI scorecards
→
- A LangGraph agent that turns business goals into structured KPI weight scorecards, with validation and fallback logic.
07SnapCal
Food photo calorie estimates
→
- A Flask app that recognizes food in photos with Inception v2 and estimates calories from nutrition data.
08House Price Prediction
EDA + four regression models
→
- Data cleaning, EDA, and feature engineering, then four regression models compared on MSE and R².
Published work
Topic Modeling Enhancement using Summaries Generated by LLM Models
2025 International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA) · Antalya, Türkiye · August 2025
ACDSA
2025
What I use
Analytics & BI
SQL, Python (Pandas, NumPy), Power BI (DAX), Matplotlib, Seaborn, R
Data quality & governance
Informatica, Alation, Dataiku
Machine learning & AI
Scikit-Learn, PyTorch, TensorFlow, LLMs, RAG, NLP, BERTopic, LangGraph
Data platforms & cloud
AWS (S3, Bedrock, OpenSearch, Amplify), Google Cloud, Hadoop, Spark, Hive, MongoDB
Apps & tooling
Flask, FastAPI, Python Dash, GitHub, Bitbucket, Jira, CI/CD
Honours Bachelor of Computer Science (Data Analytics)
Sheridan College, Oakville · GPA 3.55/4
Peer Mentor, Sheridan College
Promoted to lead 20 mentors, running sessions for 50+ students each week.
Let's talk
I'm open to data analyst and data science roles in the Toronto area or remote. The fastest way to reach me is email.
bhagyeshpatel0703@gmail.com