What should we price? • Where should inventory move? • What should we recommend?
Vehicle Pricing ↑ ~18% profit per car 80% of pricing automated Built an automated vehicle pricing system for production pricing decisions. |
Supply Chain Optimization ~7% potential cost savings 10K–15K transfers/month Built an ML decision-support system for stock-transfer optimization. |
E-commerce Recommendations ↑ ~30% email-attributed sales Developed a recommendation system for personalized email recommendations. |
Lately I've been working with LLMs: agentic RAG with LangGraph and Llama 3.1, multi-agent workflows with CrewAI, and evaluation harnesses using RAGAS.
Toolkit: Python, SQL, PyTorch, Spark, Kafka, Databricks, Snowflake, dbt, MLflow, AWS, Azure
M.S. Computer Science, RIT. Research on ML surrogates for thermal prediction published at ASME FEDSM 2026.
ML Surrogate for Heat Sink Cooling 
25 OpenFOAM CFD runs → 1.9M mesh samples → PyTorch surrogate → full 3D temperature field in 40 ms
Thermal design is bottlenecked by simulation: one CFD solve of a pin-fin heat sink takes ~15 minutes, so sweeping 1,000 geometries costs ~250 CPU-hours. A feed-forward network trained on 25 simulations predicts the entire temperature field in 40 ms — a 22,500× speedup — at MAE 0.021 K and R² 0.999997 across 383,138 held-out points. Accepted to ASME FEDSM 2026.
Randomised holdout → uplift models → budget-constrained targeting → Streamlit app
Retail customers never click cancel, they just stop coming back, so churn has to be inferred rather than observed. Using a campaign's randomised control group to measure what an offer actually caused, the finding is that the ranking flips depending on what you count: at a 5% budget, targeting by likely response returns 147 extra purchases per 1,000 customers against 32 for targeting by churn risk, but the keenest responders spend the least, so weighting response by value gives ₽1.24M incremental revenue against ₽0.27M. Neither "who is leaving" nor "who responds" is enough alone.
PriceWise — C2B Vehicle Pricing
Used car data → CatBoost on log price → pricing rules → FastAPI → React ops console
Not a notebook but a pricing product: a user enters vehicle details and the system predicts market value, recommends an acquisition offer, estimates margin after reconditioning, flags quotes needing approval, and logs whether each recommendation was accepted, rejected, or overridden — closing the loop between the model and the people using it.
→ All Data Science & ML projects
AI Powered Financial Document Analysis
Financial PDFs → FAISS + BM25 hybrid retrieval → LangGraph agent → Llama 3.1
An AI system that reads financial reports and answers questions about the numbers inside them, reviewing and correcting its own answers before responding. It scored 41.5% on FinanceBench — more than double the 19% scored by GPT 4 Turbo.
ML & Modeling
GenAI & LLMs
Cloud & DevOps
- ASME FEDSM 2026 — Machine learning for instant prediction of spatial temperature variations in heat sinks for computer chip cooling
- ✍️ Blog posts on Medium
📫 Open to Data Science roles roles · reetu.thimmaiah@gmail.com