Applied ML & Forecasting Lead for Freight MNC-Pune

Responsibilities

● Build and maintain the forecasting arms behind one common interface: seasonal baselines, statistical models,

regularised linear models, LightGBM panel models with conformal calibration, and a pre-trained challenger

(Chronos-type foundation model).

● Own walk-forward backtesting, the sealed lockbox, prediction intervals, abstention rules and the router that picks an arm per market segment; hierarchical reconciliation from lane to corridor to index.

● Run the pre-registered signal-validation loop: screening of economic hypotheses, multiple-testing (false-

discovery-rate) control, bootstrap and placebo confirmation, and the out-of-sample gate that decides whether a signal enters the models.

● Graph: evaluate graph features through the same backtest fabric as every other feature; specify what the graph layer must deliver for forecasting; keep the door open for graph neural networks once the gates are met.

● Produce reproducible evaluation reports (skill vs. naive anchor, coverage, Diebold-Mariano) verifiable from a repository tag.

Must have

● Time-series, panel and multi-horizon forecasting in production (not only notebooks); LightGBM and regularised regression; temporal feature engineering.

● Rigorous evaluation: walk-forward and temporal cross-validation, leakage detection, point-in-time (as-of) correctness, calibration, conformal prediction, prediction intervals, model comparison.

● Statistics for signal discovery: lead/lag analysis, Granger-type tests, multiple testing and false-discovery-rate control, bootstrap methods.

● Strong Python, SQL, Pandas/Polars, experiment tracking (MLflow or comparable); containers, APIs and cloud basics.

● Can instruct a coding agent to implement a new forecasting arm with backtests, leakage checks and baseline comparison — and then judge whether the result is statistically correct.

Nice to have

Hierarchical forecasting and reconciliation; foundation models for time series; graph machine learning (PyTorch Geometric, heterogeneous graph models); basic Cypher; exposure to logistics, freight, commodities or financial time series.

Looking for Someone who has shipped forecasting models to production, not only in notebooks, and whose strength is statistics and time series rather than LLM applications.

Common requirement

AI-native engineering is mandatory. Daily, productive use of Claude Code, Codex CLI, Cursor or comparable agents; breaks work into

agent-ready specs and steers parallel tasks; reads, challenges and corrects AI-generated code; tests, CI/CD and evals as guardrails

(build → test → review → fix → merge → verify); fluent in Git, pull requests and code review; maintains CLAUDE.md / AGENTS.md;Independent, remote, fluent English. Agent experience alone is not enough — LangChain/RAG/prompt-engineering profiles are not a fit.