Research & Projects
Building explainable, safe, and trustworthy AI that ships into real systems.
Research Statement
My work starts from one commitment: a claim an AI system makes about itself should be checkable by someone else.
That commitment came out of a practical problem. Chain-of-Thought reasoning made language models look inspectable, and then measurement showed the traces verbalize the actual decision process only 25 to 39% of the time, degrading further exactly where the task is hard. In CT-SAFR I took the position that this is not a problem to be solved by making explanations more honest, but one to be routed around: safety layers that do not depend on the reasoning being faithful at all. In TRACE I took the other half, building the audit requirement into the decision architecture so an incident can be reconstructed from sensor evidence rather than guessed at. In my work on soft decision trees the same commitment appears as model choice, where five interpretable parameters matched what seventeen opaque splits achieved, and where the comparison only counted because a 5×2cv F-test said the difference was real.
Three questions follow from this.
First, what can be verified about a reasoning system without trusting its explanation? I want to characterize which safety properties survive an unfaithful trace, and which quietly do not.
Second, what does auditability cost? TRACE reports sub-millisecond overhead in simulation, which is a beginning, not an answer. The real question is the frontier: how much traceability can a system carry before it changes the behavior being traced.
Third, how do we know a safety claim is true? Safety results are reported the way accuracy results were reported twenty years ago, as single numbers on single runs. A 94% detection rate on 500 samples and a 94% rate on 50 are not the same claim, and our field mostly writes them the same way. This is also why the teaching platform I built puts statistical evidence at its center rather than at the end.
Focus Areas
Explainable & Trustworthy AI
Frameworks that make AI decisions traceable and auditable, designed around emerging governance requirements such as the EU AI Act and ISO 13482.
Safe Chain-of-Thought Reasoning
Defense-in-depth verification across structural, physical, semantic, and interpretability layers that detects hallucinated or unsafe reasoning before a robot acts.
Interpretable Machine Learning
Neural-tree ensembles and other hybrid architectures for robust, explainable prediction under sensor noise and missing data.
Applied LLM Systems (PropTech / LegalTech)
Retrieval-grounded, compliance-aware LLM systems with citation-backed explanations and jurisdiction-aware safety guardrails.
Selected Projects
-
2026–
safecot-ros2: Slip-Aware State Estimation and Safety Monitoring
A measurement rig for wheel slip, ported from a ROS 1 Noetic and Gazebo Classic stack and containerized so every run is reproducible, with 22 unit tests and scripted experiment sweeps. Its first result did not survive: a rig fault found later had the robot resting on its front edge at 13.3 degrees, so the chassis dragged and manufactured the slip being measured. With the robot level, slip was re-established on a controlled low-friction ramp, where the robot travels 2.79 m while its odometry reports 9.65 m, against a high-friction control ramp that climbs the same slope with 0.34% error. Fusing an IMU with wheel odometry through
robot_localizationrecovers 9 mm of that 6.85 m error. Every claim is measured against the simplest alternative before it is reported. -
2020–
Vardenus: AI-driven PropTech & LegalTech Platform
Leading the technical vision for an AI-driven rental platform that connects landlords, tenants, and contractors: an LLM and retrieval mediation module giving citation-grounded, jurisdiction-aware guidance; compliance-aware safety guardrails; and scalable cloud and MLOps infrastructure. Tokenized ownership and escrow-automation workflows were designed and are the subject of a U.S. patent application. Recognized with a 2025 Global Recognition Award (Real Estate Technology).
-
2025–
OrthoAtlas: Clinical RAG Assistant for Orthodontics
A retrieval-augmented assistant for orthodontics and dentistry that grounds every answer in clinical sources to minimize hallucination, with a case-analysis workflow and a clinician-in-the-loop design: the tool supports the decision, the clinician makes it. Built on Render, Supabase, Qdrant vector search, and Voyage AI embeddings, and available in Turkish and English. The system was evaluated under the name AlimGPT in a blinded comparison against GPT-4o, Gemini, and Llama, published in Dentistry Journal.
-
2026
CT-SAFR: Safe & Interpretable Chain-of-Thought for Robots
A defense-in-depth, multi-layered verification framework addressing the faithfulness problem in Chain-of-Thought-enabled autonomous robots. It reported 94.2% hallucination detection and 96.4% combined detection across its four layers at sub-500 ms latency in a warehouse-robot case study, and is covered by a pending U.S. provisional patent.
-
2026
TRACE: Transparent Reasoning Architecture for Autonomous Robots
A model-agnostic, four-layer decision framework that traces every autonomous action back to sensor evidence via documented causal chains, with high evidence-traceability and decision-reconstructability across simulated decision cycles.
-
2026
ML Academy: an interactive curriculum that makes students prove what they learn
A free curriculum of 123 interactive lessons on machine learning, deep learning, and large language models, running entirely in the browser with no installation. Existing visual resources show models; this one asks the student to predict before seeing, to feel the need for a tool before it is introduced, and to prove a claim by executing the algorithm. Every algorithm engine (CART, random forests, boosting, MLP with backpropagation, SVM, soft decision trees, PCA, BPE, word2vec) was written from scratch, and a verification script re-derives every number quoted in every lesson.
-
2025
Explainable Neural Trees for Remaining-Useful-Life Prediction
An explainable neural-tree ensemble with "sensor-dropout augmentation" for robust remaining-useful-life prediction under sensor noise and missing data. It achieves strong accuracy with gradient-based feature-importance explanations and large gains in fault tolerance.
Open-Source Software
-
Web
ml-academy
The engine and the content behind mltraining.org. Content is stored as data rather than code, so a new lesson is an object, not a component. Ships with a verification script that re-derives every numeric claim, checks lesson structure, and renders all visualizations against a mock canvas.
-
PyPI
reasongate
A self-hostable gate that inspects text going into and out of an LLM and returns an explainable allow, flag, or block decision with a machine-readable audit record for every call. It de-obfuscates common evasions before matching, scans retrieved context and tool output for indirect injection, and checks output for leaked secrets and a planted canary. Every block carries a reason. Install with
pip install reasongate. -
PyPI
-
GitHub
cognitest-framework
An LLM-assisted test-automation framework exploring cognitive approaches to software testing, related to the verification ideas behind CT-SAFR.
-
GitHub
-
GitHub
