About the Role
We’re looking for a postgraduate or doctoral intern to tackle an emerging problem at the intersection of AI, legal technology, and commercial strategy.
Most AI contract analysis today evaluates whether a system can recognize clauses, spot deviations, or flag legal risks. But in real enterprise negotiations—especially in sectors like Oil & Gas, Energy, Infrastructure, and Aerospace—contracts are more than legal documents. They encode commercial relationships, risk allocation, incentives, and strategic intent.
The question we’re asking is:
Can an AI system truly understand the commercial dynamics of a contract, not just its legal provisions?
And more importantly:
How should we evaluate that understanding?
This is not a traditional NLP benchmark problem. It’s a research challenge that sits at the intersection of LLM evaluation, knowledge representation, ontology design, and real-world decision support.
What You’ll Work On
You’ll explore open research problems including:
-
Commercial relationship evaluation: How to assess whether a system has correctly inferred the nature of a commercial relationship—risk transfer, strategic partnership, cost-plus delivery—rather than just identifying parties and obligations.
-
Clause–objective alignment: How to connect individual clauses to each party’s commercial objectives, constraints, and incentives, and evaluate whether that alignment is correctly understood.
-
Detection of uncovered commercial issues: How to evaluate whether a system can identify important commercial questions the agreement does not address—gaps that could lead to value leakage or execution friction.
-
Opportunity discovery vs. risk detection: How to build evaluation frameworks that reward identification of value-creating opportunities, not just risk flags.
-
Intent inference and recommendation quality: How to assess whether a system’s negotiation recommendations are appropriate to the relationship, context, and stage of negotiation.
We’re not looking for a single large model as the answer. Instead, we’re interested in composable, interpretable, and evaluation-aware architectures. Two technical directions are particularly central to this work:
Knowledge Graphs for Commercial Relationships
We believe contract understanding requires moving beyond flat text representations. We’re exploring knowledge graphs that capture not only parties, obligations, and clauses, but also:
-
Dependencies across clauses (e.g., how a liability cap interacts with indemnity obligations)
-
Commercial objectives and incentives attached to each party
-
Economic asymmetry and risk allocation patterns
-
Operational constraints that exist outside the contract text but influence its execution
These graphs would serve as reasoning substrates, not just indexes. Key research questions include:
-
What schema best represents commercial intent alongside legal structure?
-
How can we construct these graphs from contract text with high fidelity?
-
How can graph traversal enable inference about commercial outcomes that aren’t explicitly stated?
Knowledge Distillation: LLMs as Teachers for SLMs
General-purpose LLMs are powerful but often impractical for enterprise contract workflows—they’re expensive, opaque, and hard to control. We’re interested in whether specialized Small Language Models (SLMs) can retain commercial reasoning ability at a fraction of the cost, with tighter interpretability.
This opens several research directions:
-
Distilling reasoning chains: Using LLMs to generate step-by-step commercial reasoning traces (e.g., “Clause X limits supplier liability → this shifts operational risk to buyer → given buyer’s margin constraints, this creates exposure → recommended counterproposal is Y”), then training SLMs to reproduce these reasoning paths.
-
Synthetic data generation for commercial scenarios: Using LLMs to generate diverse negotiation scenarios, redlining examples, and edge cases that would be difficult to source from real contracts at scale.
-
Contrastive learning between legal correctness and commercial effectiveness: Training models to distinguish between a clause that is legally acceptable and one that is commercially favorable given specific objectives.
-
Preference learning from accepted vs. rejected recommendations: Using real-world negotiation outcomes—what negotiators accepted versus pushed back on—as training signal for SLMs.
Key research questions include:
-
What is the minimal viable model size for retaining useful commercial reasoning?
-
Which commercial abstractions are most important to preserve during distillation?
-
Can SLMs outperform LLMs on intent- and opportunity-focused tasks when evaluated with commercial criteria rather than purely legal ones?
Who You Are
You’re a postgraduate or doctoral student with a background in:
-
NLP / LLMs
-
AI evaluation and benchmarking
-
Knowledge graphs or structured knowledge representation
-
Ontology design
-
Or a related field with a strong interest in real-world, high-stakes applications
You’re excited by problems that don’t have ready-made benchmarks. You care about whether AI actually helps people make better decisions—not just whether it achieves a higher F1 score. You’re comfortable thinking about systems that combine structured knowledge, reasoning traces, and evaluation-aware design.
What You’ll Gain
-
Hands-on exposure to real enterprise contract workflows in regulated industries like Oil & Gas, Energy, and Infrastructure
-
Experience designing evaluation frameworks from first principles, not just applying existing benchmarks
-
Deep exploration of knowledge graphs for representing commercial intent and contract structure in a way that enables reasoning beyond text
-
Hands-on work with knowledge distillation—using LLMs to generate reasoning traces and synthetic data to train smaller, more interpretable SLMs
-
The chance to shape how AI is assessed in legal and commercial domains—a space that’s rapidly evolving and underexplored
-
Mentorship from a team working at the intersection of AI and supply chain technology
How to Apply
Please share:
-
A brief summary of your research interests
-
Any relevant experience (projects, papers, or open-source work—especially if related to LLM evaluation, knowledge graphs, distillation, or contract/legal AI)
-
Why this problem interests you
-
Share the resume via michael@dandilion.ai and let me know.