Commotion builds an AI operating system for large enterprises, and at the centre of it is a context graph: a live model of an organisation’s entities, the relationships between them, and the decisions taken across them, assembled from the systems that already run the business.

Most enterprise AI stops at retrieval over documents. We think the layer that matters next is structural — resolved entities, explicit relationships, and a record of every decision and the reasoning behind it. A dashboard tells you what happened. A context graph lets an agent work out what to do about it, and show its working.

We build this inside our clients’ own environments, where the data is real, messy and consequential. We are backed by Tata Communications.

About this role

Most of this job is finding signal in data that is dirty, incomplete and contradictory, and then building models that hold up on it. The modelling is the smaller half. If you are looking for a role that is mostly about large language models, this is not it.

What you’ll do

- Translate a client’s business problem into a machine learning problem. Sit with the people who own the outcome, work out what decision is being made and what it costs to get wrong, then decide what the model predicts, what the label is, and how the prediction reaches the decision. Say so when the answer is not a model at all.
- Set the business metric and the model metric separately, and keep them connected. A lift in AUC that moves nothing operationally is a failed project here.
- Profile the data before modelling it. Work out what is trustworthy, what is abandoned, what is quietly wrong. Establish a baseline before proposing anything learned.
- Engineer features from warehouse and graph data, and keep them consistent between training and serving.
- Build the models: classification and regression, propensity and churn, lookalike and audience expansion, ranking and recommendation,



unsupervised clustering and segmentation, anomaly and exception detection, time-series forecasting, uplift and bandits.
- Apply graph algorithms where structure is the signal: community detection, centrality, link prediction, node embeddings.
- Extract structure from documents using existing OCR and parsing engines, and own the accuracy of the output.
- Deploy, then own what follows: retraining cadence, validation, champion and challenger, drift monitoring, and the call on when a model gets pulled.
- Own entity resolution: blocking and candidate generation, match scoring, threshold design, clustering records into canonical entities, and re-resolution when a record changes. Precision and recall here are your numbers, and every agent, score and client-facing answer inherits them.
- Decide which records a model can honestly be trained on, and make the case for every exclusion.
- Set the bar for how modelling work is done here, and bring newer engineers up to it.

What we’re looking for

- 4 to 9 years shipping models into production and owning them after launch.
- You can take an ambiguous business ask and frame it as an ML problem: the decision, the objective, the label, the unit of prediction, the evaluation, and the fallback when the model is not confident. This is the single most important thing we are hiring for.
- Deep applied statistics and data analysis. You reach for a distribution, a residual plot or a cohort split before you reach for a model.
- Breadth across the classical toolkit — supervised learning, clustering, forecasting, ranking, anomaly detection.



We would rather have someone competent across all of these than excellent at one.
- scikit-learn, XGBoost or LightGBM, and a forecasting library. Solid SQL, Python and Spark at volume.
- Validation design under temporal structure: leakage, calibration, class imbalance, drift. You can explain why a model that looks strong offline fails in production.
- Experiment design and measurement.
- The judgment to tell a data problem from a model problem, and to say when the data cannot support the question.
- Entity resolution or record linkage in production using an existing library or MDM tool: blocking strategies, probabilistic and ML-based matching, threshold design, clustering, and the human review loop. You know which of a false merge and a false split costs more, and why.
- Enough seniority to hold your position with a client stakeholder who wants a different answer.

How we work

- We are not asking you to invent a matcher or train a parser from scratch. We are asking you to know what to reach for, where it breaks, and how you found that out. Bring opinions about tools you have run in production, including ones you would never use again.
- Every engineer here works with Claude and agentic coding tools daily, for exploration, transformation code, test data and analysis scaffolding. Our delivery pace assumes it. Be ready to describe how these tools changed your workflow and where you have learned not to trust them.
- This is client-facing work on client sites in India. Expect travel, and expect to sit with the client’s own data owners.

Nice to have

- Graph algorithms in production, or features derived from a graph.
- OCR and document extraction at volume.
- Record linkage, deduplication or data reconciliation.
- Causal inference or uplift modelling.
- MLflow or equivalent.
- Domain exposure in manufacturing, FMCG, insurance or banking.

📌 Senior Machine Learning Engineer (India)
🏢 Commotion
📍 India