6+ years turning data into pricing, supply chain, and product decisions
What should we price? • Where should inventory move? • What should we recommend? • And did it actually work?
Vehicle Pricing ↑ ~18% profit per car 80% of pricing automated Automated pricing for a German automobile client, with the profit lift measured through randomized A/B tests against manual pricing. |
Supply Chain Optimization ~4.5% potential cost savings 10K–15K transfers/month ML decision support for stock-transfer sourcing, saving procurement admins 1.5 hours a day. |
E-commerce Experimentation Email share of sales: 3% → 7% 10+ A/B tests per quarter Designed and analyzed email experiments, and A/B-tested a cart-abandonment recommender that contributed to ~30% quarter-over-quarter growth in email-attributed sales. |
Randomized 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 randomized control group to measure what an offer actually caused, 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 (figures in the dataset's currency, Russian rubles). Neither "who is leaving" nor "who responds" is enough on its own.
→ 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.
I also build multi-agent workflows with CrewAI and evaluation harnesses with RAGAS.
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 across 383,138 held-out points.
Publications
- ASME FEDSM 2026 — Machine learning for instant prediction of spatial temperature variations in heat sinks for computer chip cooling (lead author)
- Towards utilizing machine learning and computational fluid dynamics in the classroom for high heat dissipation
- ✍️ Blog posts on Medium
M.S. Computer Science, Rochester Institute of Technology
Experimentation & Analytics
ML & Modeling
Data Platforms
GenAI & LLMs
Cloud & DevOps
📫 Open to Product Data Science & Data Science roles · reetu.thimmaiah@gmail.com