1 / 4 · About

About

I am a researcher drawn to the mathematics of how models learn. The question I now want to study is what models actually learn, and how to make that learning reliable.

Now
Since January 2024, starting in the third year of my undergrad, I have been an AI researcher at Vela Partners, in a research group that publishes with collaborators at the University of Oxford. Our problem is predicting which early-stage startups will become unicorns. Fewer than 2% do, the data is sparse and noisy, and the outcome can look like pure luck. I build and maintain the group's dataset and started think-reason-learn, our open-source library for LLM-native models. I have co-authored eight papers and am a named inventor on a US patent application. I am most interested in the mathematical problems: normalizing inputs so a model is not fooled by them, scoring rules fairly when positives are rare, and helping models learn the right signals cheaply while they keep learning from new data.
Background
I studied Electrical & Electronic Engineering at the University of Mines and Technology, Ghana, graduating with First Class honors. I was always most excited by problems where a learned model has to act in the real world. On UMaT's Drillbotics team I led the software for two entries that placed at the top of their design options: a digital-twin planner for directional drilling, and a model that detects dangerous pressure kicks from a virtual rig's real-time sensor data and triggers automated well control, with a human override, because a wrong call on a rig is costly. I also led the research for that second entry. The kick-detection model was built to connect to physical rigs through the D-WIS interface. The work led to two SPE conference papers.

Next: Research

Research

Direction I want to understand what deep learning and reinforcement learning models learn and why, and use that to make them learn faster, generalize further, and behave more reliably, especially where data is scarce and mistakes are costly. Three questions interest me most:
  • What do models represent internally, and can we read, test or correct it?
  • How can a model focus on learning the right signals at a low cost, and keep learning from new data without forgetting what it knows?
  • How do these methods hold up where data is scarce and mistakes are costly: in physical systems (control, robotics, dynamical systems), in healthcare, and in other high-stakes decisions?

Interpretable decision-making with LLMs

Instead of asking a language model for an answer, the group asks it for a decision structure: a tree split, a yes/no question, a probabilistic rule, a piece of executable code. The structure is then fitted and validated statistically, so every prediction has a path a person can read.

Inventor on the GPTree patent application; implemented the Random Rule Forest ensemble and much of this line in think-reason-learn.

GPTree · GPT-HTree · Random Rule Forest · LLM-AR (automated reasoning) · Verifiable reasoning · CoFEE (feature discovery)

Benchmarks and data

Rare-outcome prediction needs clean, leak-resistant data. I build and maintain the dataset of startups and founders the group's models are trained on, and contributed to VCBench, the first standardized benchmark for founder-success prediction.

VCBench · Rare-event prediction

Learning for physical systems

I led the software for UMaT's 2023 Drillbotics entry, a digital-twin planner for directional drilling, and the research and software for its 2024 well-control entry, an LSTM over real-time sensor data from a virtual rig simulator that detects kicks and triggers automated well control with human override. The D-WIS well-monitoring paper comes out of that work. My undergraduate capstone was a smart-metering analytics system for electricity distribution. This is where my interest in learning for physical systems comes from.

Drillbotics · D-WIS well monitoring

Software

think-reason-learn

An open-source, LLM-native machine-learning framework: scikit-learn-style estimators whose splits, rules, and questions are proposed by a language model, then fitted and validated like any other model. It brings the group's LLM-built models, among them GPTree and Random Rule Forest, under one interface. MIT licensed.

I started the library and remain a core contributor.

Documentation · GitHub

Next: Publications

Publications

Selected

  1. VCBench: Benchmarking LLMs in Venture Capital

    Chen, Ternasky, Kwesi, Griffin, Yin, Salifu, Amoaba, Mu, Alican, Ihlamur · Intelligent Computing, Springer · 2026

    The first benchmark for founder-success prediction: 9,000 anonymized profiles with re-identification tests, and nine LLMs evaluated against human investors. The benchmark is built from the startup and founder data I maintain.

  2. Random Rule Forest (RRF): Interpretable and Manageable Ensembles of LLM-Generated Questions for Predicting Success from Unstructured Data

    Griffin, Yin, Vidaurre, Koyluoglu, Ternasky, Alican, Ihlamur · arXiv:2505.24622 · 2025

    An LLM writes simple yes/no questions as weak learners; a unit-weight vote turns them into an auditable scorecard that is robust when positives are scarce. I implemented the ensemble in think-reason-learn.

  3. Learning What to Ask and When to Stop: Cost-Aware Sequential Founder Evaluation

    Ye, Alican, Griffin, Yin, Ihlamur · Under review, ICAIF 2026

    A reinforcement-learning policy decides which attributes to query next and when to stop, keeping most of the accuracy at about half the information cost. Runs on the same data. I suggested the features and wrote the code for leakage-safe extraction, and caught errors in the action-space normalization and the policy loss.

All papers

Author lists as published. Preprints are marked with their arXiv identifier.

  1. 2026

  2. CoFEE: Reasoning Control for LLM-Based Feature Discovery

    Maximilian Westermann, Ben Griffin, Aaron Ontoyin Yin, Zakari Salifu, Yagiz Ihlamur, Kelvin Amoaba, Joseph Ternasky, Fuat Alican, Yigit Ihlamur

    arXiv:2604.21584 · April 2026

    Treats feature discovery as a reasoning problem and enforces cognitive behaviors on the LLM so the features it proposes are predictive rather than leaky proxies.

  3. From Stochastic Answers to Verifiable Reasoning: Interpretable Decision-Making with LLM-Generated Code

    Anirudh Jaidev Mahesh, Ben Griffin, Fuat Alican, Joseph Ternasky, Zakari Salifu, Kelvin Amoaba, Yagiz Ihlamur, Aaron Ontoyin Yin, Aikins Laryea, Afriyie Samuel, Yigit Ihlamur

    arXiv:2603.13287 · March 2026

    One LLM call generates executable decision logic that runs deterministically over the whole dataset, replacing per-sample queries with reproducible, statistically validated rules.

  4. Integrating Machine Learning with Domain Expertise for Smarter Real-Time Well Monitoring Using D-WIS Interface

    Joel Sekyi Mensah, Aaron Ontoyin Yin, Gabriel Kowfie, Janet Intuah, Fedra Mensah Martha, Precious Segoe, Richard Amorin

    SPE Ghana Biennial International Conference & Exhibition · May 2026

    Combines a machine-learning layer with driller expertise on the D-WIS data interface for real-time well monitoring.

  5. VCBench: Benchmarking LLMs in Venture Capital

    Rick Chen, Joseph Ternasky, Afriyie Samuel Kwesi, Ben Griffin, Aaron Ontoyin Yin, Zakari Salifu, Kelvin Amoaba, Xianling Mu, Fuat Alican, Yigit Ihlamur

    In Intelligent Computing (Computing Conference 2026, London), Lecture Notes in Networks and Systems, Springer, pp. 167–187 · June 2026 · free version: arXiv:2509.14448

    Nine thousand anonymized founder profiles, adversarial re-identification tests, and nine LLMs benchmarked against the market index and top-tier investors.

  6. 2025

  7. LLM-AR: LLM-powered Automated Reasoning Framework

    Rick Chen, Joseph Ternasky, Aaron Ontoyin Yin, Xianling Mu, Fuat Alican, Yigit Ihlamur

    arXiv:2510.22034 · October 2025

    Distils LLM-generated heuristics into probabilistic rules executed by ProbLog, with an iterative policy-evolution loop, so every prediction exposes its decision path.

  8. From Limited Data to Rare-event Prediction: LLM-powered Feature Engineering and Multi-model Learning in Venture Capital

    Mihir Kumar, Aaron Ontoyin Yin, Zakari Salifu, Kelvin Amoaba, Afriyie Kwesi Samuel, Fuat Alican, Yigit Ihlamur

    arXiv:2509.08140 · September 2025

    LLM-extracted features from unstructured data feed a layered ensemble of classical models that estimates and then thresholds the likelihood of a rare outcome.

  9. Random Rule Forest (RRF): Interpretable and Manageable Ensembles of LLM-Generated Questions for Predicting Success from Unstructured Data

    Ben Griffin, Aaron Ontoyin Yin, Diego Vidaurre, Ugur Koyluoglu, Joseph Ternasky, Fuat Alican, Yigit Ihlamur

    arXiv:2505.24622 · May 2025

    An LLM writes yes/no questions as weak learners; a unit-weight vote turns them into an auditable scorecard that is robust when positives are scarce.

  10. Dual Approach in Autonomous Directional Drilling: Innovations with Drillbotics 1.5 Inch Automated RSS and Virtual Rig Platforms

    C. Soilemezidis, J. S. Mensah, W. Hollstein, K. Al Maasarani, A. Alkhawaja, F. Jamali, G. Dörffler, N. Hölzner, K. A. Jarbouh, D. Ustaoglu, A. T. Alukkal, L. A. Y. Gudhane, P. Jaeger, E. Feldmann, C. A. P. Carvajal, A. O. Yin, L. Sinkuu, J. Quayson, F. B. Amoah, H. S. Yahaya, K. A. Owusu, C. K. N. Dongoyo, A. Owusu, R. Amorin, F. Florence

    SPE/IADC International Drilling Conference and Exhibition · March 2025 · DOI 10.2118/223656-MS

    Joint paper from the Drillbotics 2024 teams on a physical 1.5-inch automated rotary-steerable rig and a virtual rig platform for directional drilling.

  11. GPT-HTree: A Decision Tree Framework Integrating Hierarchical Clustering and Large Language Models for Explainable Classification

    Te Pei, Fuat Alican, Aaron Ontoyin Yin, Yigit Ihlamur

    arXiv:2501.13743 · January 2025

    Clusters individuals first, fits a decision tree per cluster, and uses an LLM to write readable descriptions of each segment.

  12. 2024

  13. GPTree: Towards Explainable Decision-Making via LLM-powered Decision Trees

    Sichao Xiong, Yigit Ihlamur, Fuat Alican, Aaron Ontoyin Yin

    arXiv:2411.08257 · November 2024

    A decision tree whose splits are proposed and evaluated by an LLM from a single task prompt, with an expert-in-the-loop step to refine decision paths.

Under review

  1. Learning What to Ask and When to Stop: Cost-Aware Sequential Founder Evaluation

    Yuhang Ye, Fuat Alican, Ben Griffin, Aaron Ontoyin Yin, Yigit Ihlamur

    Under review, ICAIF 2026

    A reinforcement-learning policy decides which attributes to query next and when to stop, keeping most of the accuracy at about half the information cost.

Patent applications

  1. Large language model (LLM) powered decision trees

    Yigit Ihlamur, Sichao Xiong, Fuat Alican, Aaron Ontoyin Yin

    US patent application 19/357,996 · 2026

Next: Experience

Experience

  1. Jan 2024 – present

    AI Researcher, Vela Partners

    San Francisco, remote from Ghana

    • Build and maintain the group's dataset of tens of thousands of startups and their founders, collected from public and licensed sources, on which the group's models are trained.
    • Built the founder-success prediction models in production use at the firm, with precision about ten times the 1.9% base rate; now building their second version.
    • Started think-reason-learn, the group's open-source LLM-native machine-learning framework, and remain a core contributor.
    • Co-authored eight papers with the group and am an inventor on the GPTree patent application.

  2. 2023 – 2024

    Drillbotics, UMaT team

    Global student competition run by the SPE Drilling Systems Advancement Technical Section

    • Led software development for the 2023 entry, and the research and software for the new 2024 well-control entry.
    • 2023, virtual rig, directional drilling: built the digital-twin planner; first among directional-drilling entries and second overall in the virtual-rig category. UMaT News · MyJoyOnline · Drillbotics.com
    • 2024, virtual rig, directional drilling: helped another UMaT team improve the planner, which placed first in the virtual-rig category.
    • 2024, virtual rig, well control (new for UMaT that year): led the research and software and built an LSTM over real-time sensor data from a virtual rig simulator that detects kicks and triggers automated well control with human override; first among well-control entries and second overall in the virtual-rig category.
    • Joint paper with the other teams at the SPE/IADC Drilling Conference, 2025. Paper

Education & Honors

Education

  • B.Sc. Electrical & Electronic Engineering, University of Mines and Technology, Tarkwa, Ghana. Jan 2022 – Sep 2025. First Class, 3.91/4.0.
  • Capstone: design of a smart-metering analytics system for electricity distribution.

Honors

  • Best Student in Software Engineering, UMaT Excellence Awards, 2024
  • IEEEXtreme programming competition, 1st among Ghanaian teams, 2023

Leadership

  • Vice President, Aaenics UMaT Robotics Club, 2024/25. Taught members electronics and embedded programming: microcontrollers, Arduino, and Python.

End · Back to start