Shamin
Chokshi
I build agentic AI systems, RAG pipelines, and ML at production scale — turning messy data and language into products that save money and ship fast.
I turn ambiguous problems into agents that act, and data into decisions.
I'm a Generative AI Data Scientist at Bright Horizons with a Master's in Data Architecture & Management from Northeastern. My work lives at the intersection of large language models, autonomous agents, and applied machine learning.
From RAG-powered counseling chatbots to multi-agent fraud detection that prevented $10M in erroneous payouts, I care about systems that are reliable, measurable, and genuinely useful — not demos. I ship end-to-end: data pipelines, training, evaluation, and CI/CD.
Northeastern GenAI Hackathon winner. Builder of things people actually use.
Where I've built.
Generative AI Data Scientist
- ›Built a RAG agentic chatbot for personalized student-loan & tuition-reimbursement counseling, cutting support requests and operational cost 9%.
- ›Shipped a multi-agent system that retrieves and visualizes client data from complex reports via natural language — manual lookup time down 90%.
- ›Built a LangGraph + Azure OpenAI agent detecting tampered grade sheets, preventing $10M in fraudulent tuition reimbursements.
- ›Forecasted unique EdAssist users at 96% accuracy with a SARIMA time-series model for revenue & budget planning.
- ›Deployed a course recommendation engine on Azure ML, lifting certification registrations 13%.
- ›Built an end-to-end MLOps pipeline (Jenkins / GitHub Actions) with a precision/recall/F1 evaluation framework.
Business Data Analyst Intern
- ›Built a multimodal hybrid-search chatbot (LangChain · AWS · Pinecone, voice + text), cutting operational cost 30%.
- ›Orchestrated KPIs from an EDW via SQL/ETL & BI tooling, improving feature-refinement accuracy 40%.
- ›Built Power BI & Tableau dashboards driving real-time decisions, +35% monitoring efficiency.
Software Data Engineer · Finance
- ›Modeled 650K financial transactions for portfolio optimization & risk management.
- ›Engineered a network-analysis framework (DAGs, community detection) mapping transaction pathways across 10K client profiles.
- ›Designed pipelines processing 1M+ financial data points, raising data quality and analysis efficiency.
Software Data Engineer · Healthcare
- ›Feature engineering & dimensionality reduction on 10K+ EHR datasets, +25% model performance.
- ›Deployed a patient-readmission ML pipeline (scikit-learn · TensorFlow) with A/B testing, −11% readmissions.
- ›Ran clustering & ensemble models identifying high-risk segments, +15% recovery-prediction accuracy.
How I learned to build.
Northeastern University
M.S. in Data Architecture & Management
Graduate study across data systems, machine learning, and applied AI.
Vellore Institute of Technology
B.Tech in Computer Science & Engineering
Foundations in algorithms, systems, and software engineering.
The stack I build with.
Languages
- Python · PyTorch
- TensorFlow · Keras
- SQL · Pandas
- C / C++ · JavaScript
GenAI / Agents
- LangChain · LangGraph
- Multi-Agent · MCP
- RAG · Retrieval
- Anthropic Claude · OpenAI
- Google ADK · RLHF
Cloud / Platforms
- Azure AI · Azure ML
- AWS Bedrock
- Databricks · Snowflake
- Azure Data Factory
- Vector DBs
ML / MLOps
- Model Building · NLP
- Forecasting · A/B Testing
- LLMOps · Deployment
- CI/CD · GPU
- Multimodality
Databases
- PostgreSQL · MySQL
- MS SQL · Oracle
- MongoDB · Cassandra
- DynamoDB · SQLite
Analytics / BI
- Power BI · Tableau
- Data Analytics
- ETL · EDW
- Network Analysis
Things I shipped for fun and real users.
less-tokens
An MIT-licensed Python library that compresses LLM prompts before they hit the API — cutting 30–40% of tokens with no GPU, no extra model, running in milliseconds. compress_structured shrinks instructions while leaving JSON schemas and rules untouched, and compare scores quality across six metrics so you prove the output holds. Backed by a study on ~18,000 LLM completions. Compare before you compress.
pip install less-tokens↗Handmade AI
An AI image pipeline that turns any photo into a numbered paint-by-numbers stencil in under 30 seconds — KMeans color segmentation, Canny edges, bilateral filtering. A free, open-source replacement for $15–30/image services.
Tweak AI
A 4-agent autonomous pipeline (LangGraph + GPT-4) for keyword extraction, experience refinement, project optimization & skills tailoring — cutting resume tailoring from 45 min to under 60 seconds. Shipped as a Chrome extension with LinkedIn auto-detection at 80% less cost than competitors.