Co-founder building products, growth, and AI systems at INHAUS Coffee. This is where I keep the technical proof of that range: nine projects spanning AI/ML, product, growth, marketing, and startup finance, each with a real evaluation and an honestly-reported result — including the ones where a simpler baseline beat the fancier approach.
AI / ML systems
- should-cost-copilot — flagship project. An RFQ/BOM review copilot combining a parametric cost engine, ML quote-range prediction, and a retrieval-augmented DFM guidance assistant into one workflow. FastAPI + Streamlit, both installed dependencies of the two projects below, not copy-pasted logic.
- dfm-guideline-rag — hybrid (dense + BM25) retrieval over an original DFM/GD&T knowledge base, fused with Reciprocal Rank Fusion, evaluated with Recall@K/MRR and refusal accuracy against deliberately out-of-scope questions.
- cost-estimation-engine — a transparent parametric should-cost model plus an ML layer for quote prediction, uncertainty intervals, SHAP explanations, and outlier detection.
- sku-demand-forecasting — walk-forward-validated demand forecasting comparing naive/ETS/SARIMA baselines against a global gradient-boosted model, across 20 SKUs with different demand personalities.
Product & growth
- product-growth-analytics — activation funnel, cohort retention, a segment/channel-controlled feature-adoption model, a North Star Metric, and a proper A/B-testing engine that reports an honestly underpowered test rather than a false positive.
- product-strategy-toolkit — RICE vs. WSJF prioritization compared head-to-head on the same backlog (they agree on only 2 of 5 top items), a Kano survey simulated and correctly classified, and a capacity-aware roadmap scheduler — plus a full PRD and an evidence-linked positioning case study.
Marketing & business analytics
- marketing-mix-model — adstock + saturation MMM and Markov-chain multi-touch attribution on synthetic data with known ground truth. Honest finding: MMM's channel-level decomposition is unstable and shrinkage-biased, while path-based Markov attribution recovers ground truth almost exactly.
- d2c-growth-analytics — cohort retention, RFM segmentation, and channel CAC/LTV, built as real SQL (CTEs, window functions) over a DuckDB warehouse, not one pandas query pretending to be analytics.
- startup-unit-economics — cohort LTV/CAC, a Monte Carlo cash-runway simulation, and a cap-table/exit-waterfall model, tied into a real fundraising memo: strong unit economics (9.8x LTV:CAC) don't prevent a 74.8% chance of running out of cash within 24 months.
Currently working with: Python · SQL · scikit-learn · DuckDB · FastAPI · Streamlit · sentence-transformers · statsmodels
Background: IIT Jodhpur. Based in Bengaluru.