Kemal Özkırşehirli / Projects

Projects

Research, engineering, writing, and selected experiments.

No projects match that filter.

Principal Investigator — MeshAnyOrder for Life Sciences - Apr 2026 – Present - Leading a 7-member independent research collaboration developing MeshAnyOrder, an order-agnostic autoregressive transformer for point-cloud-conditioned 3D mesh generation. The model represents faces as quantized tokens and predicts unvisited adjacent faces from arbitrary traversal seeds. Current work emphasizes 3D rotary positional encoding, triangle/quad tokenization, topology-aware validity constraints, frontier-parallel decoding, local mesh completion/remeshing, and ablations across traversal order, masking, manifoldness, watertightness, latency, memory, and high-resolution scaling.

Geometric Deep Learning - 3D Meshes - Autoregressive Models - Topology - Life Sciences

Principal Investigator — TBXT / Brachyury Small-Molecule Discovery - Apr 2026 – Present - Leading an 11-person chordoma-focused computational hit-identification project targeting PDB 6F59 chain A / TBXT G177D site F. The workflow compresses 2,274 prior-art compounds, 737 raw analogs, 503 filtered analogs, 30,000 BRICS recombinations, and 67 QSAR-pass proposals into a 570-compound novelty-filtered pool, then applies docking, GNINA CNN rescoring, RF/XGBoost QSAR, Boltz-2 co-folding, MMGBSA/FEP scaffolding, T-box paralog selectivity, Rowan affinity analysis, PAINS/sourceability filters, and Bash/HPC automation.

CADD - Docking - QSAR - TBXT - Chordoma - Drug Discovery

PRISM AI Safety Fellow — Protein Foundation Model Red-Teaming - May 2026 – Present - Selected from 800+ applicants for PRISM 2026 to develop adversarial evaluation methods for protein foundation models. The work tests biological plausibility, sequence–structure–function consistency, structural validity, developability, uncertainty calibration, claimed mechanisms, and failure behavior, and builds safety-bounded benchmarks and taxonomies for high-stakes protein-design workflows.

AI Safety - Protein Foundation Models - Evaluation - Benchmarking

ChromoGen-Engine V2 — Conditional Diffusion for 3D Chromatin - May 2026 – Present - Developing a conditioned 3D genome generation and evaluation pipeline for single cells. The work frames chromatin structure as an intermediate physical layer between DNA sequence, regulatory state, cell identity, and chemical/regulatory perturbation; builds preprocessing and conditioning machinery for sequence windows, perturbation labels, and regulatory readouts; and evaluates coordinate ensembles with contact maps, P(s), polymer/geometric realism, condition separation, calibration, leakage checks, and attribution diagnostics.

Diffusion Models - Genomics - Chromatin - Single-Cell - Multi-Omics

A*STAR V2M-Engine — Autoimmune Target Discovery - May 2026 – Present - Building a provenance-first, leakage-controlled variant-to-mechanism evidence graph for autoimmune/IBD target discovery. The workflow links disease, locus, variant or QTL/colocalization evidence, effector genes, cell contexts, mechanisms, and target-control tiers; starts with transparent classical scoring and source-pinned evidence; and leaves Perturb-seq, sequence-effect, single-cell foundation-model, and targetability modules as downstream extensions after the evidence graph is inspectable.

Target Discovery - Single-Cell - Genetics - Autoimmune Disease - Drug Discovery

Pedal AI — Chemistry Research-Implementation and Backend Engineering - Jul–Dec 2025 - Designed Python/LangGraph and FastAPI infrastructure connecting chemistry agents with molecular-prediction engines, supporting stateful orchestration, model/data evaluation, and inspectable retrosynthesis workflows. Built chemist-facing UI/UX around Coley Lab / ASKCOS-style route search, reaction-level provenance, and decision-ready synthesis planning.

LangGraph - FastAPI - Cheminformatics - Agentic AI - Backend Engineering

VeriQSM + QSMBench — Artifact-Grounded Scientific Verification - v0.3 research release - VeriQSM separates nominal workflow completion from scientifically verified success. The framework converts scientific intent into typed, auditable workflows for quantum chemistry and statistical mechanics, then applies static checks, physics verifiers, reference-style calculations, provenance contracts, bounded recovery/refusal policies, and cost/reproducibility accounting to detect false success in agentic scientific workflows.

Scientific Agents - Quantum Chemistry - Statistical Mechanics - Verification - Benchmarking

Deep Reinforcement Learning for Antibody–Antigen Interactions - Research scaffold - Research scaffold for risk-constrained, structure-guided antibody sequence optimization. The implementation combines ESM-2-compatible antigen embeddings, structure-informed cross-attention for antibody CDR generation, OAS/SAbDab/IEDB-style curation, supervised warm starts, selective PPO, uncertainty penalties, developability screens, contamination-aware splits, evidence-lineage controls, and audit-ready evaluation scaffolding.

Reinforcement Learning - Antibodies - Protein Models - ESM-2 - Scientific Software

EVEdesign — Protein Design Algorithm Engineering - Apr 2026 – Present - Contributing to EVEdesign's open-source, method-independent protein-design platform as part of a 21-person, 18-institution, 8-country collaboration. Work includes uncertainty-aware candidate selection, sequence–structure–function objective integration, protein-language-model/evolutionary priors, agent-based tool orchestration, and scalable lab-in-the-loop biosequence design.

Protein Design - Open Source - Uncertainty - Biosequence Design

Kadanoff-GNN-RG — ML-Augmented Renormalization Group - Alpha research software - Research framework for ML-augmented hierarchical-lattice renormalization-group analysis. The system combines direct continuum RG operators, finite-spin distribution comparisons, symmetry/gauge-aware features, typed-edge graph neural networks, Wasserstein/MMD-style distribution distances, calibrated phase classification, empirical RG-flow reconstruction, and uncertainty-aware phase-topology analysis.

Statistical Mechanics - Renormalization Group - GNNs - Monte Carlo - Scientific Software

Kupcinet–Getz Reaction–Diffusion AI - v0.3 benchmark and software release - Scientific-computing framework for nonlinear biochemical reaction–diffusion systems, oscillators, traveling waves, and synthetic morphogenesis. It combines deterministic ODE solvers, chemical-Langevin dynamics, Gillespie SSA, spatial RDME, convergence audits, solver-agreement checks, calibration, abstention, and evidence manifests to study when numerical representation changes inferred dynamics.

Reaction–Diffusion - Scientific AI - Stochastic Simulation - PINNs - Verification

Sepal AI / Mercor — Chemistry LLM Evaluation - Sept 2024 – Feb 2025 - Developed model-training, testing, and evaluation data for graduate-level chemistry reasoning, including retrosynthesis, reaction mechanisms, and method selection. Designed grading rubrics, chemical-validity constraints, instructional data, prompt-engineering workflows, error analysis, and structured fine-tuning/evaluation formats.

LLM Evaluation - Chemistry - Scientific Data - Prompt Engineering

DFT → kMC — Auditable Multiscale Reaction Kinetics - v0.1 implementation scaffold - Mechanism-aware decision-stability workflow linking Gaussian/ORCA transition-state thermochemistry to Eyring rates, deterministic ODEs, exact finite-state CME solutions, and Gillespie/SSA stochastic kinetics. The framework evaluates whether solvent, pathway, or product rankings remain stable under declared energetic, mechanistic, and model-form uncertainty while preserving provenance and reportability.

DFT - Kinetic Monte Carlo - Eyring Theory - Computational Chemistry - Reproducibility

Advanced Organic Synthesis — Buchwald Lab - Dec 2025 – May 2026 - Supported development of CuH-catalyst chemistry for selective asymmetric methylation and alkylation of vinyl boronate esters, used 1H-NMR and chiral-HPLC to relate catalyst structure to activity, expanded substrate scope, and scaled synthesis from 0.1 mmol to multi-gram quantities.

Organic Chemistry - Synthesis - Catalysis - NMR - HPLC

MIT 6.1200[J] Teaching Fellow and Grader - Jan 2026 – Present - Selected after receiving an A+ and ranking in the top 1% of MIT EECS's largest foundational theoretical-computer-science course; supports a 250+ student cohort through recitations, office hours, proof-intensive grading, midterms, and finals in logic, graph theory, recurrences, asymptotics, and cryptography.

Teaching - Discrete Mathematics - Proofs - MIT EECS

ClubChem - Present - Participation in MIT's undergraduate chemistry community.

MIT - Chemistry - Community