Lead Machine Learning Engineer (Core Modelling) — 7+ yrs

Indpro AB; Bengaluru; On-site; INR 1,800,000 - 3,000,000; Full time

Job summary

Indpro AB in Bengaluru, India seeks an experienced ML professional to own problems end to end, frame problems for forecasting, classification, clustering, or anomaly tasks, and prove them in production. You will build and train classical, probabilistic, and deep models with state-of-the-art tools, emphasizing rigorous evaluation and leakage prevention.

The role requires fluency in Python, experience with scikit-learn, PyTorch or TensorFlow, and strong data wrangling skills.

Qualifications

• Genuine core-ML fundamentals.
• Fluency in Python and the modern ML stack.
• Experience with model evaluation and metrics.
• Strong data-wrangling and SQL skills.

Responsibilities

• Own problems end to end, from framing to production.
• Build and train models — classical, probabilistic, and deep.
• Choose evaluation metrics that fit the problem, and defend them.
• Engineer features and handle messy reality.
• Prevent, detect, and interrogate leakage.

Skills

Genuine core-ML fundamentals.

Fluency in Python

Cross-validation

Metrics selection

Model evaluation

Tools

scikit-learn

PyTorch

TensorFlow

Job description

We build and ship machine learning models that make real decisions in production — forecasting, anomaly detection, ranking, classification, and probabilistic systems that customers and revenue depend on. This is a core modeling role. You will spend most of your time framing problems in modeling terms, choosing and training the right models, and proving they work under real-world conditions — noisy data, distribution shift, class imbalance, and the constant threat of leakage.

We are deliberate about one thing: this role is about understanding models, not just wiring them together. We use LLMs and retrieval where they earn their place, and applied GenAI experience is a genuine plus. But if your experience is primarily prompt engineering, RAG pipelines, and vector-database orchestration, this is not the right role — and we'll both save time by being upfront about that now.

If the most interesting thing you did last year was pick a loss function, catch a subtle leak that was inflating your metrics, or prove a simpler model beat a fancier one, we want to talk to you.

What you'll do

• Own problems end to end, from framing to production. Take an ambiguous business problem, decide whether it's a forecasting / classification / clustering / anomaly / ranking / probabilistic problem, and justify that framing. Wrong framing is the most expensive mistake in ML, and we expect you to get it right.
• Build and train models — classical, probabilistic, and deep. Depending on the problem: tree ensembles (XGBoost / LightGBM / CatBoost), time-series models (ARIMA/SARIMA, Prophet, state-space, TCN, and modern sequence models), probabilistic models (HMM, GMM, Bayesian methods), and deep networks (CNNs, RNN/LSTM, transformers) where they genuinely outperform. Fine-tune models properly (LoRA/PEFT, full fine-tunes, or training from scratch) when that's the right tool — not as a euphemism for calling an API.
• Choose evaluation metrics that fit the problem, and defend them. PR-AUC vs ROC-AUC under imbalance, WMAPE vs MAPE for intermittent demand, calibration when probabilities matter. Report the metric that tells the truth, not the one that flatters the model.
• Engineer features and handle messy reality. Class imbalance (weighting, resampling, threshold tuning), missing data (principled imputation), multicollinearity, and temporal structure. Know when a feature is quietly leaking the target.
• Prevent, detect, and interrogate leakage. Time-based and grouped splits, walk-forward / rolling-origin validation, and a healthy suspicion of any result that looks too good. We consider "I caught a leak" a badge of honor, not an embarrassment.
• Validate rigorously and report honestly. Cross-validation strategy appropriate to the data, held-out and locked test sets, bootstrap intervals where useful, and clear separation of durable gains from overfitting. Negative results and killed models are expected and valued.
• Ship and operate what you build. Package, version, deploy, and monitor models — batch and real-time. Track drift (data / concept), automate retraining, and build in rollback. You don't have to be a full MLOps specialist, but your models should survive contact with production.
• Use GenAI where it earns its place. Some of our systems combine classical models with retrieval or LLM components. You'll build these when they're the right answer — with the same evaluation rigor you apply to everything else.
• Communicate and collaborate. Explain modeling choices to engineers, product, and non-technical stakeholders. Tie model quality to business outcomes without letting the business framing paper over weak modeling.

What we're looking for (required)

• Genuine core-ML fundamentals. You can explain, from first principles: the bias–variance trade-off, why a model is overfitting and how you'd know, what a given loss function optimizes, how regularization works, and why your chosen validation scheme is sound for this data.
• A track record of models you actually built and evaluated — with the architecture, the data, the metric, and the before/after numbers. "Implemented an ML model for various use cases" is not a track record. We want the specifics.
• Fluency in Python and the modern ML stack: scikit-learn, and at least one of PyTorch / TensorFlow used for real training (not just inference). Solid SQL and data-wrangling (pandas / PySpark or equivalent) for large datasets.
• Sound experimental discipline: appropriate cross-validation, leakage prevention, metric selection, and the judgment to prefer a simpler model when it wins.
• The instinct to quantify. You reach for numbers by default, and you know the difference between a self-reported metric and an externally validated one.
• Strong communication and intellectual honesty — you can defend a modeling decision and, just as importantly, admit when a result doesn't hold up.

Nice to have (preferred, not required)

• Deep-learning depth: architectures built or meaningfully modified from scratch, training-dynamics debugging, or published/research work.
• Probabilistic and Bayesian modeling; time-series beyond the standard toolkit (state-space, hierarchical, foundation models for forecasting).
• Genuine fine-tuning experience (LoRA / PEFT / full) with a clear eval story — and the ability to say precisely what it accomplished over prompting.
• Production MLOps maturity: CI/CD for models, drift detection (PSI / Evidently), automated retraining, canary releases, and observability.
• Applied GenAI / RAG experience — welcome as a complement to modeling depth, not a substitute for it.
• Kaggle, open-source ML contributions, or peer-reviewed publications.
• Domain experience in [your domains — e.g., forecasting, fraud/risk, healthcare, retail, manufacturing].

What this role is not

To keep us both efficient, this role is not a fit if your experience is centered on:

• Prompt engineering and LLM orchestration as the primary skill.
• Building RAG pipelines (chunking, embedding, vector search, retrieval) without underlying modeling work.
• Agentic-workflow wiring (LangChain / LangGraph / CrewAI) as the core of your portfolio.
• Integrating third-party model APIs without training or evaluating models yourself.