Principal Scientist (Neuro-Symbolic AI)

Company: Blue Fire AI

Team: Decision-Engine Research & Development

Location: Singapore (hybrid; some collaboration across Japan, UK and US time zones), Employment Pass will be sponsored

Level: Senior / Principal Research Scientist

Reports to: CEO

Requirement: PhD, plus hands-on ability to write production-quality code and design original algorithms

About Blue Fire AI

Blue Fire AI is a technology-led asset management and investment intelligence firm. Our core platform, Emmalyn, is a neuro-symbolic fundamental-analysis engine. Emmalyn operates in two commercial modes: Risk Analyst, producing decision-ready outputs for investment managers and exclusion lists; and Investment Manager, where the engine drives allocation through a risk-alpha overlay.

We are not building a black-box predictor. Every risk call we ship must carry an auditable chain of evidence: which entities, which relations, which filings, which events, and what would have had to be different for the call to flip. That requirement — machine reasoning that is both learned and explainable — is why this role is crucial.

The Role

You will design and build the next version of the reasoning core of Emmalyn: the layer that combines learned representations over company, executive, board, supply-chain and market graphs with explicit symbolic constraints, causal structure, and counterfactual evaluation. Concretely, you will turn methods from the current neuro-symbolic and causal-reasoning methods into production algorithms that generate, explain, and stress-test equity risk and alpha alphas across a global universe of thousands of listed companies.

This is a research role with shipping obligations. You will publish internally, defend your methodology to portfolio decision-makers and institutional clients, and own the code that runs in the pipeline with implementation support from the AI development technologists.

What You Will Do

Expected || Reasoning architecture

• Design the hybrid learning-and-reasoning core: differentiable logic layers (e.g. Logic Tensor Networks / Real Logic-style grounding of first-order signatures onto data), energy-based logical inference, and constraint injection into neural models so domain rules — accounting identities, governance rules, index methodology, exclusion policy — are enforced rather than hoped for.
• Build System 1 / System 2 style architectures: fast learned screening over the full universe, with slow, deliberate symbolic verification and meta-cognitive arbitration deciding when a case must escalate to the reasoning layer.
• Develop knowledge representation for our knowledge graphs, and reason over them with graph neural networks used as neuro-symbolic machinery (relational reasoning, constraint satisfaction, structured inference) rather than as generic embedders.

Expected || Causal and counterfactual machinery

• Implement counterfactual reasoning on top of structural causal models — identification, interventional vs. counterfactual layers, canonical/normalized representations that separate unfalsifiable counterfactual assumptions from testable interventional constraints — and make explicit which of our claims live on which rung of the causal ladder.
• Build counterfactual evaluation for our own learning systems: off-policy / counterfactual estimation of "what would this alpha set have produced had we changed the rule," importance-weighted estimators with confidence intervals, and equilibrium-aware reasoning about a deployed system interacting with its environment.
• Develop anomaly attribution for multivariate time series using counterfactual replacement of variable subsets — moving us from "an anomaly fired" to "these three drivers explain it, and here is the evidence."
• Build evidence subgraph extraction for financial risk: Granger-causality-style meta-path attribution, edge-type-aware generators, feature masking, and joint counterfactual-plus-factual objectives so each risk flag ships with a minimal sufficient and necessary subgraph as its explanation.

Aspirational || Abduction and uncertainty

• Design abductive inference for event and narrative reasoning — given an observed outcome and prior context, generate and rank the most plausible explanatory hypotheses, including backprop-based or search-based decoding that conditions on both past and future context.
• Represent non-probabilistic uncertainty where probability is the wrong tool: logic programming with three-valued/Kleene semantics, default and non-monotonic reasoning, normative and I/O logic for policy and mandate constraints.
• Build calibration and predictive-uncertainty layers, and explicit tests for out-of-distribution failure and for cases where theoretical guarantees break.

Productization

• Own the algorithms and the reasoning trace that makes each defensible to an institutional client reviewer.
• Work with the alpha research team on evaluation: CAR-based backtests, hit rates, cohort cuts, borrow-cost-adjusted net edge, capacity analysis, and honest median-versus-mean reporting.
• Write clean, tested, reproducible Python; contribute to pipeline design; leave documentation a colleague can rebuild from.

Required Qualifications

• PhD in Computer Science, Artificial Intelligence, Machine Learning, Statistics, Applied Mathematics, Operations Research, Computational Logic, Econometrics, or a closely related field, with a dissertation or publication record in reasoning, causality, logic, or structured/relational learning.
• Demonstrated ability to design algorithms from first principles — not only apply libraries. We will ask you to derive, prove or bound something on a whiteboard and then implement it.
• Strong engineering ability: expert Python; fluency in at least one deep learning framework (PyTorch, JAX, or TensorFlow); comfort with SQL/Aurora and large tabular/time-series data; version control, testing, and reproducible experiment discipline.
• Depth in at least three, and working literacy across most, of the following:
• Neuro-symbolic / neural-symbolic computing: differentiable logic, fuzzy/many-valued semantics, knowledge injection, rule extraction, energy-based logical inference (e.g. RBM/Logical Boltzmann-style systems)
• Causal inference: SCMs, do-calculus, identification, counterfactual identification, canonical representations, distributional regression
• Counterfactual learning systems: off-policy evaluation, importance sampling and clipped estimators, confidence bounds, counterfactual policy improvement
• Graph representation learning: GNNs, heterogeneous and knowledge graphs, meta-paths, relational and combinatorial reasoning, graph explainability
• Abductive and non-monotonic reasoning; logic programming; answer set programming; probabilistic logic
• Multivariate time-series modelling, anomaly detection, and attribution
• Constrained optimization and its integration with learning (MILP, robust/stochastic optimization)
• Explainability as a design goal — you have built systems where the explanation was a first-class output with its own evaluation metrics.
• Ability to read a research paper on Monday and have a defensible prototype by Friday.
• Clear technical writing and the ability to present methodology to non-AI experts, including investment professionals and clients.

Preferred Qualifications

• Publications at NeurIPS, ICML, ICLR, AAAI, IJCAI, KDD, ACL, UAI, CLeaR, or equivalent venues in neuro-symbolic AI, causality, or graph reasoning.
• Experience in finance: equity fundamentals, credit or governance risk, factor/alpha research, event studies, index methodology, or regulatory filings (10-K/20-F/annual reports and disclosure regimes).
• Experience with NLP over long financial documents: event extraction, entity linking, transcript and disclosure analysis, and schema-bound information extraction.
• Familiarity with neuro-symbolic toolchains (LTN, DeepProbLog, Scallop, PyReason, NeurASP, ProbLog, Answer Set solvers) and graph stacks (PyG, DGL, Neo4j).
• Knowledge-graph construction at scale, including entity resolution across executives, boards, subsidiaries and dual listings.
• Experience taking research into production: cloud pipelines (AWS), batch and streaming orchestration, and monitoring for model drift.
• Meta-cognition, self-monitoring, or agent-architecture research — an acknowledged gap in the current neuro-symbolic literature and an area we intend to lead in.

How We Evaluate Candidates

1. Technical screen — your research: what you built, what you proved, what broke.
2. Algorithm design exercise — design a reasoning layer for a stated financial risk problem, on a whiteboard, with your assumptions made explicit.
3. Code exercise — implement a small differentiable-logic or counterfactual-attribution component and evaluate it honestly, including failure modes.
4. Paper deep-dive — pick one paper from the neuro-symbolic or causal literature and teach it to us, including its weaknesses and what you would do differently.
5. Investment-facing conversation — explain your method to someone who allocates capital and cares about auditability.

What We Offer

• Ownership of the reasoning core of a live investment engine, with real capital and institutional clients downstream of your work — not a research sandbox.
• Proprietary datasets: a multi-year global company, executive and board knowledge graph, event and downgrade histories, and alpha outcome data with forward return windows.
• A small, senior, low-bureaucracy team; direct access to the founder; freedom to publish where it does not compromise IP.
• Competitive base, performance-linked compensation tied to engine and product outcomes, and support for conference attendance and continued research.