Lavi Sahu — Index of Work & Thinking

Which node, if it fails, hurts most — and for how long? Time-to-Recover and Time-to-Survive stress testing in the Simchi-Levi tradition. Identifies single points of failure in multi-tier supply networks.

Node-criticality scoring via supply chain network graph analysis. Deterministic TTR/TTS computation across tiered supplier dependencies. Synthetic multi-echelon data, all labelled as such.

Python, NetworkX, pandas. Deterministic seeded data for full reproducibility.

GitHub — supply-chain-resilience

Is our forecast actually adding value, and where is it worst? Forecast Value Added (Gilliland) combined with ADI/CV² demand segmentation (Syntetos–Boylan–Croston) to separate signal from noise.

ADI/CV² quadrant classification of SKU-level demand. FVA decomposition across forecasting horizons and method overrides. Separates intermittent, lumpy, smooth, and erratic patterns.

Python, pandas, synthetic demand data. Reproducible from seed — same inputs, same answer.

GitHub — demand-planning-diagnostics

Base vs upside vs constrained — what do we commit, and what does it cost? S&OP/IBP scenario reconciliation with rough-cut capacity analysis. Bridging demand consensus with supply feasibility.

Multi-scenario demand-supply balancing (Wallace / Oliver Wight framework). Rough-cut capacity checks against constrained supply options. Gap analysis with economic cost of constrained decisions.

SAP IBP, o9, spreadsheet models. 20+ programs delivered across pharma cold chain, FMCG, and energy.

GitHub — sop-integrated-planning

Notes on automation applied to real planning workflows — not theory, not hype. What works when you plug scripts into demand forecasting, S&OP reconciliation, and KPI reporting. Practical, reproducible, grounded.

Practical write-ups with reproducible code. Every claim grounded in a live process, every dataset synthetic and labelled as such. No assertions without computation.

Operational planning notes from 14 years across pharma cold chain, FMCG, and energy. What the textbooks don't cover: handoff politics, data quality anti-patterns, stakeholder alignment mechanics, the gap between model elegance and shop-floor reality.

First-principles reasoning applied to messy realities. Bounded claims only — if the evidence is thin, the note says so. Judgment over code craft.

Where AI actually helps supply-chain planning — and where it doesn't. LLM-assisted report drafting, anomaly detection heuristics, prompt-engineered data QA. Skepticism-first, honest about limitations.

Skepticism-first: every AI claim tested against a non-AI baseline. Honest about compute costs, latency budgets, and the irreducibility of domain judgment. "The forecast is a decision, not a prophecy."

SAP IBP — demand, supply, response & supply planning. o9 — AI/ML demand sensing and granularity management. Kinaxis RapidResponse — concurrent planning and scenario simulation. OMP — granularity-first S&OP. SAP APO — legacy DP/SNP (still live in many organisations).

14 years, 5 platforms, 3 industries deep (pharma cold chain, FMCG, energy), 20+ programs delivered. Functional configuration and delivery leadership. Every number computed, not asserted.

Good analysis is bounded by what the data supports. Overreaching models damage trust faster than underperforming ones. Say what you know, flag what you don't. Underpromise, overdeliver.

Every model ships with confidence intervals, known limitations, and conditions for invalidation. No silent extrapolation. If the claim is thin, the deliverable says so.

Every number computed, not asserted. Synthetic data labelled synthetic. If you can't reproduce it from seed, it's opinion dressed up as analysis. Magic numbers are forbidden.

All repos use deterministic seeds. Same inputs, same answer. No hardcoded constants without provenance. Every dataset ships with a data dictionary.

"The forecast is a decision, not a prophecy." Forecasts exist to enable action, not to predict the future. The moment you stop asking "what do we do with this?" the forecast becomes cargo — expensive, inert, and taking up space.

Every forecast deliverable includes a recommended action section. If nobody acts on it, it doesn't ship. The value of a forecast is measured in decisions changed, not MAPE improved.

Reproducible — seed everything; same inputs, same answer. From first principles, not from best practices. Every analysis traceable to its assumptions. If the assumptions change, the analysis changes.

Config-driven pipelines with version-controlled assumptions. Seed management as a first-class concern. Every parameter file ships alongside the code it drives.