Advancing AI Reliability: An Interview with Joseph Oladokun on Path-Constrained Retrieval and Real-World Machine Learning

By: Jessica Williams

Joseph Oladokun is a Data Scientist whose research and work is addressing one of artificial intelligence's most pressing challenges: making LLM agents reliable for real-world reasoning tasks. His recent publication on Path-Constrained Retrieval introduces a novel approach to semantic search that could transform how enterprises deploy AI systems. With a track record spanning healthcare, fintech, and enterprise software across both Nigeria and Silicon Valley, Joseph represents a new generation of AI researchers focused on bridging the gap between theoretical innovation and practical deployment.

Joseph, your recent paper "Path-Constrained Retrieval: A Structural Approach to Reliable LLM Agent Reasoning Through Graph-Scoped Semantic Search" tackles a fundamental problem in AI. What motivated this research?

Joseph Oladokun: The motivation came from a frustration I'd repeatedly observed in production environments: LLM agents were unreliable for tasks that require genuine reasoning. They would hallucinate connections between pieces of information, fail to maintain logical consistency, or struggle with temporal relationships in data. Traditional retrieval-augmented generation (RAG) approaches treated retrieval as a flat search problem, but real-world reasoning requires understanding structural relationships between information.

Path-Constrained Retrieval (PCR) fundamentally reimagines how agents interact with knowledge graphs. Instead of simple semantic similarity searches, PCR constrains retrieval paths based on the actual structural relationships in the data—whether that's causal chains, temporal sequences, or hierarchical dependencies. This ensures that when an LLM agent retrieves information to reason with, it's pulling contextually appropriate information that respects the underlying logic of the domain.

Can you give us a concrete example of where this makes a difference?

Joseph: Consider a healthcare scenario—something I worked on during my time at Helium Health in Lagos, Nigeria. When a doctor is reviewing a patient's history to make a diagnosis, they're not just looking for similar symptoms. They're tracing causal chains: this symptom appeared after this medication was prescribed, which was given because of this prior condition. That's a constrained path through the knowledge graph of the patient's history.

With traditional RAG systems, an LLM might retrieve superficially similar cases but miss the critical temporal and causal relationships. PCR ensures the retrieval follows the logical structure—the paths that actually matter for sound medical reasoning. The same principle applies in financial risk assessment, legal case analysis, or any domain where structural relationships are crucial.

You've published this research on arXiv and open-sourced the implementation on GitHub. What's been the response from the research community?

Joseph: The response has been encouraging. I've had researchers from several institutions reach out to discuss applications of PCR to their specific domains. What's particularly validating is hearing from practitioners—engineers at companies building AI agents—who recognize this addresses a real pain point they're experiencing in production systems.

The open-source release was intentional. I believe AI advances fastest when researchers can build on each other's work. By making both the paper and code available, I'm hoping to see PCR extended and improved by others tackling similar challenges.

Your career spans multiple continents and industries—Nigeria, Lithuania, and now Silicon Valley, with experience in healthcare, fintech, and enterprise software. How has this shaped your approach to AI research?

Joseph: That cross-industry, cross-continental perspective has been invaluable. At Helium Health, I built an NLP system that converted unstructured doctor notes into structured medical codes—ICD-10 and RxNorm format. That experience taught me that AI systems need to be reliable and explainable, especially in high-stakes domains like healthcare.

With RataFX, the fintech company I co-founded in Lagos, we scaled to 5,000 users and processed $450,000 in gross merchandise value within four months. That taught me about building systems that work under real-world constraints—regulatory uncertainty, infrastructure limitations, and the need for user trust.

At Asana, I'm applying these lessons at scale. I've built uplift models, opportunity scoring systems, and LLM pipelines for analyzing sales transcripts. Each role reinforced the same lesson: the value of AI isn't in benchmark performance—it's in solving actual problems reliably.

You mentioned your medical NLP work at Helium Health achieved impressive results. Can you elaborate on that system?

Joseph: At Helium Health, we faced a challenge common in healthcare: doctors write notes in natural language, but for billing, analysis, and interoperability, those notes need to be converted into standardized codes—ICD-10 for diagnoses and RxNorm for medications. This was being done manually, which was time-consuming and error-prone.

I developed an NLP model using ClinicalBERT with PyTorch, We achieved 86% accuracy, which was sufficient for the system to handle the majority of cases automatically, with only edge cases requiring human review. The real impact was reducing manual mapping time by 95%—freeing healthcare workers to focus on patient care rather than administrative coding.

I also built the MLOps infrastructure to deploy this reliably. In healthcare, you can't afford system downtime or model failures, so the infrastructure work was just as critical as the model itself.

Let's talk about your journey from Nigeria to Silicon Valley. What challenges did you face, and what drove you to pursue AI research?

Joseph: I completed my undergraduate degree in Nigeria at the Federal University of Technology, Akure, then worked across several African tech companies. Each role showed me how AI could transform industries that had traditionally been underserved by technology.

The challenge was always resources—computational resources, access to cutting-edge research, the infrastructure that Silicon Valley companies take for granted. But constraints breed creativity. I learned to build efficient systems, to focus on practical impact rather than chasing the latest trends.

I pursued a master's degree at Iowa State University to deepen my theoretical foundation while maintaining that practical focus.

You're also working on FullGradient, a tool for fine-tuning LLMs. What gap does this fill?

Joseph: FullGradient addresses a practical barrier: fine-tuning large language models is too complex for most organizations. You need expertise in distributed computing, GPU optimization, and model architectures.

FullGradient is a tool that abstracts away that complexity. It allows teams to retrain LLM models with customized data without needing to worry about GPU management, distributed training setups, or the technical details of model optimization. The goal is to democratize access to fine-tuning.

You've also contributed to the broader AI community through writing. Tell us about your work as a technical author.

Joseph: I co-authored "JupyterLab Quick Start Guide" published by Packt, which introduced data scientists to the next generation of interactive computing environments. I also regularly write on Medium, covering topics in machine learning, data science, and AI. The goal with my writing is always to make complex concepts accessible and to share practical insights from building real systems.

Writing forces clarity of thought. It's also a way to give back to the community—I've learned so much from others' blog posts and tutorials, so contributing my own is a way of paying that forward.

Looking at the AI landscape today, where do you see the most important challenges and opportunities?

Joseph: The biggest challenge is reliability. We have incredibly powerful models, but we can't always trust them for high-stakes decisions. That's what motivated my work on Path-Constrained Retrieval—addressing the reasoning reliability problem at a fundamental level.

The opportunity is in practical deployment. There's a massive gap between what's possible in research labs and what's actually deployed in production systems. Researchers who can bridge that gap—who understand both the theoretical foundations and the practical constraints of real-world systems—can have enormous impact.

What's next for your research?

Joseph: I'm continuing to develop Path-Constrained Retrieval, exploring applications in different domains, and working to make it more efficient and accessible. I'm also interested in temporal reasoning for foundation models—helping LLMs better understand how information changes over time and how to reason about dynamic environments.

Ultimately, my goal is building AI systems that actually work in the real world, not just in academic benchmarks.

Any advice for aspiring AI researchers, particularly those from non-traditional backgrounds or underrepresented regions?

Joseph: Focus on solving real problems. The most impactful work comes from deeply understanding actual problems people face and building solutions that work in practice.

Also, don't underestimate the value of a non-traditional path. My experience across different industries and continents gave me perspectives that purely academic researchers might miss. The constraints I faced—limited resources, infrastructure challenges, diverse problem domains—forced me to think creatively and focus on what truly matters.

Finally, contribute to the community. Open-source your work, write about what you learn, help others who are earlier in their journey. The AI field advances through collective effort, and there's room for diverse perspectives and approaches.

Joseph Oladokun's work exemplifies a new generation of AI research—deeply grounded in theory but laser-focused on practical impact. His Path-Constrained Retrieval approach represents not just a technical contribution, but a philosophical shift toward building AI systems that are reliable, explainable, and ready for real-world deployment.

Connect with Joseph:

LinkedIn: linkedin.com/in/joexy/

GitHub: github.com/Godskid89

Medium: medium.com/@oladokunjoseph2

Research: arxiv.org/abs/2511.18313