Research Scientist — Decision Intelligence

You will build the layer where optimisation, machine learning and human judgment meet: interpretable decision strategies that can be fully explained , driven by learned components that a solver can execute.

Reports to: CEO, Design and Build Location: Singapore (hybrid) Level: Scientist / Senior Scientist

The mandate

Blue Fire AI converts company fundamentals, events and risk features into decision- ready investment products. Two research frontiers define this role.

• First, interpretable decision strategies: rather than choosing between white-box

heuristics and black-box ensembles, you compose them — using post-hoc explainable outputs (partial dependence, ALE, feature attribution, ensemble predictive uncertainty) as candidate inputs to sparse, auditable rule structures such as fast-and-frugal trees, with expert intervention as a designed step rather than an afterthought.

• Second, learning inside combinatorial optimisation: treating portfolio

construction, exclusion-list selection and capital-allocation problems as constrained integer programmes, and using learning to replace expensive algorithmic decisions or to discover better policies — while preserving feasibility and optimality guarantees.

What you will own

• Hybrid interpretable models. Multi-step pipelines that lift the accuracy of sparse

rule-based strategies using ensemble-derived signals, without surrendering full- model interpretability or auditability.

• Uncertainty as a first-class output. Ensemble and Bayesian variance estimates

surfaced as calibrated confidence on every risk score, score change and exclusion decision.

• Learn-to-optimise components. Learned branching, variable/cut selection, warm

starts and primal heuristics inside MILP formulations; end-to-end predict-then- optimise and decision-focused losses where the downstream objective, not predictive error, is the target.

• Problem-distribution design. Framing which family of instances the models must

generalise over, and designing the evaluation that exposes out-of-distribution failure before capital does.

• Human-in-the-loop protocol. Visualisation and intervention tooling that makes

expert overrides explicit, logged and testable — an interpretable decision-support layer, not a dashboard.

• Research-to-production handoff. Reproducible experiments, versioned features,

point-in-time correctness, and written notes that survive client and regulatory scrutiny.

Requirements

• PhD in Operations Research, Analytics, Industrial Engineering, CS or a closely

related field.

• Deep fluency in integer and stochastic/robust optimisation: MILP modelling,

duality, decomposition, and hands-on use of a commercial or open solver (Gurobi, CPLEX, HiGHS) plus modelling layers such as RSOME or JuMP.

• Demonstrated machine-learning depth — supervised learning, imitation learning

and reinforcement learning — and the judgment to know which of the three a given algorithmic decision actually needs.

• Strong Python engineering; comfort with graph neural networks or

structured/sequence models for instance representation is an advantage.

• Working understanding of explainability and interpretability as distinct concepts,

and why sparsity matters when a human must carry the rule in working memory.

What differentiates a top candidate

• You can state, unprompted, where a learned heuristic breaks a theoretical

guarantee — and what you would do about it.

• You prefer a model a human will actually use over a marginally better one they will

override silently.

• Prior exposure to finance is welcome but not required; intellectual honesty about

noisy, non-stationary data is vital.