About Yamly

AI agents now write code faster than teams can understand it or keep control of it. Yamly is building the Engineering World Model: a governed, evidence-backed model of how an engineering organisation actually works. It joins code, decisions, deploys, incidents and runtime behaviour into one record, and humans and coding agents use it before they change anything.

Every claim in the model carries evidence, a calibrated confidence and an honest causal label. A deterministic core of parsers, static analysis and a typed temporal graph supplies the structure. Your job is the learning layer on top of that core.

We’re seed-stage, and the founders have built and exited before. The first instruments already run on real code.

The role

You’ll own the science that turns structure and history into calibrated numbers. That covers change-risk scores, incident root-cause rankings, deploy-to-metric and cost attribution, and what-if simulation. You’ll set the research agenda, ship models into production at design partners, and prove the lift on their systems, measured against baselines declared in advance.

What you’ll work on

- Scoring and calibration. Build transparent risk models over symbolic features, using survival/hazard models, Bayesian shrinkage and conformal guarantees. The setting is rare labels and small org-level N.
- Causal inference on engineering data. Treat deploys, canaries and config flips as natural experiments. Methods include synthetic controls, interrupted time series, DoWhy-GCM on known graphs, and identification and refutation checks that decide what the product is allowed to claim.
- Representation learning. Train contrastive edit/diff and incident embeddings, with training pairs mined from reverts, fix→cause links and verified incident causes.
- Neurosymbolic reasoning. Build propose-and-verify loops, model-based diagnosis over the graph, and rule induction from very few examples.
- SLM specialisation.



Distil verification-filtered extractors, rankers and judges that run inside customer VPCs.
- The label factory. Produce supervision through weak supervision, mutation and fault injection, and active learning, without annotation projects.
- Evaluation. Run hindcasts on customers’ own history, benchmark against public baselines (RCAEval, SWE-bench-family, CoIR), and set kill criteria for every bet.
- Hiring and mentoring the science team as it grows.

What we’re looking for
- 8+ years in applied ML or data science, including senior or principal ownership of models that shipped and that people relied on.
- Depth in at least two of: causal inference, survival analysis or Bayesian modelling, representation learning, LLM fine-tuning or distillation, neurosymbolic/probabilistic reasoning, RL or bandits.
- Robust statistical judgment. You know when a logistic regression beats a deep net, and why, and you choose methods by data regime rather than by fashion.
- Fluent in Python and the modern ML stack. You write production-quality code, not just notebooks.
- Comfortable with messy real-world data such as git history, logs, tickets and postmortems.
- Rigour about honesty: calibration, uncertainty, and saying “cannot attribute” when the data can’t support a claim.
- Clear writing for engineers, customers and investors.

Strong plus
- Experience with software engineering data, AIOps/SRE, or developer tools.
- Publications or open-source work in causal ML, program analysis + ML, or code models.
- Experience with knowledge graphs or entity resolution over heterogeneous sources.

Why join
- Hard, open problems at the meeting point of program analysis, causality and AI agents, where the work has a direct route to production.
- Founding-team ownership and meaningful equity.
- A deterministic-first, evidence-backed culture where science is held to predeclared baselines.

To apply: email [email protected] with your CV.

Work Location: In person

📌 Principal Data Scientist (Mumbai)
🏢 Yamly Unicontext
📍 Mumbai