MOTIVATION LETTER The Michael Page Career Transition Grant exists to move talented researchers into work that reduces catastrophic risks from artificial intelligence. My research trajectory has been building toward that transition for two years, and this grant is the mechanism that makes it concrete. I am a Nigerian pharmacist and computational modeler with a B.Pharm from the University of Ibadan (CGPA 5.1/7.0, German equivalent 1.9), currently enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute / University of Potsdam. My independent research spans addiction neuroscience, protein machine learning, and dynamical-systems methods, and I am applying to use this grant to redirect my full attention to technical AI safety. The strongest evidence for my fit is my neurocascade project: a receptor-to-behavior brain-circuit simulation engine that couples pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics to behavioral-readout ODE layers. The system is calibrated for three receptor/circuit systems (mu-opioid, D2 dopamine, GABA-A), uses Bayesian MCMC calibration via PyMC, and passes 62 of 62 tests. It required me to build a multi-scale pipeline that connects molecular events to systems-level behavior, which is precisely the kind of modeling rigor that technical AI safety research demands for interpretability and control problems. The CCT model, my tripartite pharmacological framework for reward-memory encoding prevention in addiction, uses a coupled three-axis ODE model with Bayesian MCMC calibration across 14 free parameters, with all five pre-registered hypotheses confirmed. That work demonstrates I can design, pre-register, execute, and report a complex computational study honestly, including when results are negative. I have also demonstrated the capacity to pivot based on evidence. My cardiotoxicity topology study tested whether bipartite persistent homology predicts hERG cardiotoxicity from protein-ligand interface geometry. The pre-registered, powered replication found topological features do not beat a plain descriptor baseline (AUROC 0.8426 vs 0.8782). I reported that result directly. My TOPOLOGIX project then pivoted to sequence-based representations, achieving AUROC 0.804 plus or minus 0.025 on the Platinum benchmark, beating structure-based baselines while covering 100 percent of mutations versus roughly 18 percent for structure-limited tools. This pattern, form hypothesis, test rigorously, report honestly, adapt, is the core competency of AI safety research. Open Philanthropy funds work that reduces catastrophic risks from AI. My plan for the grant period is to complete the M.Sc. Digital Health coursework at HPI, which provides formal training in computational methods and AI systems, while redirecting my independent research program toward AI interpretability. Specifically, I will apply the multi-scale dynamical-systems approach from neurocascade to the problem of understanding and controlling large language model internal representations. The psyche-twin project, a typed multi-scale knowledge-graph architecture that fuses multiple evidence streams into an append-only event log with disagreement represented as explicit graph edges, is a direct foundation for work on AI self-modeling and transparency. I have the technical skills: Python, PyMC, ODE solvers, TDA, NEURON/Brian2, AlphaFold, RDKit, GROMACS, and production systems operations including self-hosted LLM serving with llama.cpp. I have the endorsements of Kent Berridge at Michigan, Samuel Gershman at Harvard, Nathaniel Daw at Princeton, and Marcelo Mattar at NYU. What I lack is the dedicated time and funding to make the transition to AI safety research full-time. This grant provides that. RESEARCH STATEMENT My research program is organized around one question: how do multi-scale dynamical systems produce behavior, and how can we model, predict, and control them? I have pursued this question in neuroscience and pharmacology, and I am now positioned to apply the same methods to AI safety. The grant period would fund a structured transition from independent computational neuroscience research to technical AI safety research, with a focus on interpretability and control of large-scale AI systems. The neurocascade project is my most direct evidence of readiness. It is a receptor-to-behavior brain-circuit simulation engine that couples four ODE layers: pharmacokinetics, receptor binding, Wilson-Cowan circuit dynamics, and behavioral readouts. The system is calibrated for three receptor/circuit systems (mu-opioid, D2 dopamine, GABA-A) using Bayesian MCMC via PyMC, with literature-elicited priors. All 62 tests pass. The circuit-layer parameters are explicitly labeled illustrative pending real behavioral-data fits, which reflects my commitment to honest reporting of validation status. This project demonstrates the exact systems-thinking and modeling rigor needed for technical AI safety research, particularly for problems involving interpretability of complex, multi-component systems. The CCT model, my Conjunctive Consolidation Threshold framework, is a tripartite pharmacological model for reward-memory encoding prevention in addiction. It uses a coupled three-axis ODE model (dopaminergic RPE, NMDAR-dependent LTP, affective contrast) solved with RK45, calibrated with Bayesian MCMC (PyMC DEMetropolisZ, 14 free parameters) against priors elicited from a 1,847-record literature screen. All five pre-registered hypotheses (H1-H5) were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. Three sole-authored preprints are in review at peer-reviewed journals (IART, PNPBP, NBR), and a co-authored paper is under review at Alcohol (Elsevier). This work demonstrates my ability to design, pre-register, execute, and report a complex computational study with rigorous statistical methods. My negative results are as important as my positive ones. The cardiotoxicity topology study tested whether bipartite persistent homology (opposition-distance metric, Ripser/GUDHI) predicts hERG cardiotoxicity from protein-ligand interface geometry. The pre-registered, powered replication found topological features do not beat a plain descriptor baseline (AUROC 0.8426 vs 0.8782). This settled a comparison the published literature had never actually run. The interface-topology-for-resistance study applied the same constructs to drug-resistance prediction and found they carry almost no signal (AUROC 0.425 and 0.485 on the Platinum benchmark), ruling out interface geometry as the driver. These results demonstrate that I can design rigorous tests of hypotheses and report negative findings without reframing them. The TOPOLOGIX project is the positive pivot from those negative results. It uses ESM-2 protein-language-model delta-embeddings plus Morgan/ECFP drug fingerprints and a Random Forest classifier to predict drug-resistance mutations from sequence alone. It achieves AUROC 0.804 plus or minus 0.025 on the Platinum benchmark (553 mutations) and 0.634 on SKEMPI 2.0, beating structure-based baselines (mCSM-lig around 0.70) while covering 100 percent of mutations versus roughly 18 percent for structure-limited tools. This project demonstrates my ability to adapt methods based on evidence and to build effective ML pipelines. The ergofluids project is a methods-validation study testing whether Koopman-operator / Dynamic Mode Decomposition methods extended with a Mori-Zwanzig memory kernel can model macromolecular drug-vehicle transport through dense, non-Newtonian tumor tissue. The pre-registered, gated validation pipeline passed synthetic-data gates, but the first real-data gate (tested against digitized published figures) did not meet its primary pre-registered criterion. I reported this directly rather than reframing it. This is a methods-validation research project, not a venture, and I make no IP or product claims. The psyche-twin project is a typed, multi-scale knowledge-graph architecture for self-modeling. Multiple independent evidence streams (LLM-derived, assessment-derived, behavioral, document-derived) fuse into one append-only event log, with disagreement between streams becoming an explicit graph edge rather than being averaged away. A first-person conversational interface sits on top of the graph, generating responses grounded in the graph's current state. This project is the most direct bridge to AI safety work on interpretability and transparency of AI systems. During the grant period, my plan is threefold. First, complete the M.Sc. Digital Health coursework at HPI/Potsdam, which provides formal training in AI systems and computational methods. Second, redirect my independent research program toward AI interpretability, specifically applying the multi-scale dynamical-systems approach from neurocascade to the problem of understanding and controlling large language model internal representations. Third, build a network in the AI safety community by attending technical workshops, engaging with Open Philanthropy-funded researchers, and publishing my methods and results in venues accessible to the AI safety community. The grant would fund dedicated research time, travel to key workshops and conferences, and the computational resources needed for LLM interpretability experiments. ESSAY: CAREER TRANSITION PLAN The Michael Page Career Transition Grant is designed for researchers who can demonstrate a meaningful shift toward AI safety work. My transition plan has three phases, each with concrete deliverables and milestones. Phase one, months one through four, is formal training and skill consolidation. I am enrolled in the M.Sc. Digital Health program at the Hasso Plattner Institute / University of Potsdam, starting Winter Semester 2026/27. The coursework covers AI systems, computational methods, and digital health infrastructure. In parallel, I will complete a structured reading and implementation program in AI interpretability, focusing on mechanistic interpretability methods (sparse autoencoders, activation patching, circuit analysis) and control methods (representation engineering, steering vectors). I will implement at least two published interpretability papers from scratch to build hands-on competence. Phase two, months five through ten, is the research pivot. I will apply the multi-scale dynamical-systems approach from neurocascade to AI interpretability. The specific project is to model large language model internal representations as coupled dynamical systems, using the same Bayesian calibration and ODE methods I have applied to brain circuits. The psyche-twin knowledge-graph architecture provides a foundation for this work, as it already handles multi-stream evidence fusion with explicit disagreement edges. The deliverable is a technical report or preprint describing the approach and initial results, submitted to a venue accessible to the AI safety community. Phase three, months eleven through twelve, is community integration and job exploration. I will attend at least two technical AI safety workshops or conferences, present my work, and build relationships with researchers in the field. I will apply for PhD programs in AI safety or computational neuroscience with an AI safety focus, and for research positions at AI safety organizations. The endorsements I already hold from Kent Berridge (Michigan), Samuel Gershman (Harvard), Nathaniel Daw (Princeton), and Marcelo Mattar (NYU) provide a foundation for these applications. The grant amount of $52,500 covers living expenses for the grant period, travel to workshops and conferences, and computational resources for LLM interpretability experiments. My current employment as National Product Manager at Synthcare (starting March 2026) provides a baseline, but the grant would allow me to reduce clinical and product-management commitments and dedicate focused time to the AI safety transition. The grant is the difference between a gradual, part-time transition and a focused, full-time one. ESSAY: RELEVANT EXPERIENCE AND SKILLS My technical skills map directly onto the requirements for technical AI safety research. I have deep experience in Python (scipy, numpy, ODE/RK45 solvers, PyMC/MCMC, pandas), which is the primary language of AI safety research. I have built and calibrated complex dynamical systems models: the CCT model uses a coupled three-axis ODE with Bayesian MCMC calibration across 14 free parameters, and neurocascade couples four ODE layers with Bayesian calibration via PyMC. This is the same methodological toolkit used in computational neuroscience approaches to AI interpretability. I have hands-on experience with protein language models and ML pipelines. TOPOLOGIX uses ESM-2 protein-language-model delta-embeddings plus Morgan/ECFP fingerprints and a Random Forest classifier, achieving AUROC 0.804 plus or minus 0.025 on the Platinum benchmark. This required me to build a complete ingest-to-analyze pipeline, manage data across formats, and rigorously evaluate model performance. I have built four independent DuckDB-based corpus/RAG pipelines across life-sciences, tech/AI/security, and social-science domains, which demonstrates my ability to handle large, heterogeneous datasets. I have production systems experience that is directly relevant to AI safety infrastructure work. I self-host local LLM serving with llama.cpp, including on-demand model swapping, on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backup/disaster-recovery. I operate these systems in production. I have experience with Nextflow/SLURM/HPC for large-scale computation, and with Supabase/Postgres for data management. My research track record demonstrates intellectual rigor and honest reporting. The cardiotoxicity topology study was a pre-registered, powered replication that found topological features do not beat a plain descriptor baseline (AUROC 0.8426 vs 0.8782). I reported this negative result directly. The ergofluids project passed synthetic-data gates but failed its first real-data gate against its primary pre-registered criterion, and I reported that failure directly rather than reframing it. This commitment to honest reporting is essential for AI safety research, where the cost of overclaiming is catastrophic. My domain expertise in pharmacology and neuroscience provides a unique angle on AI safety. I understand how complex biological systems achieve strong and control, and how they fail. The neurocascade project, which couples pharmacokinetics to receptor binding to circuit dynamics to behavior, is a working example of multi-scale modeling that AI safety researchers are trying to achieve for AI systems. The CCT model demonstrates my ability to formalize hypotheses about complex systems and test them with rigorous Bayesian methods. My professional experience includes clinical pharmacy (Ramset Pharmacy, January to March 2026), research assistance at CDDDP (NMDA/insulin docking) and GHRU-GSAR (AMR genomics, surveillance pipeline), and product management at Synthcare (starting March 2026). These roles have given me project management, collaboration, and communication skills that complement my technical abilities. CHECKLIST - [ ] Verify current eligibility criteria on the Michael Page Career Transition Grant page at openphilanthropy.org/grants/michael-page-career-transition-grant/ - [ ] Confirm the application deadline from the programme website - [ ] Verify that the M.Sc. Digital Health enrollment at HPI/Potsdam (Winter Semester 2026/27) is confirmed and documented - [ ] Confirm the status of the three preprints in review (IART, PNPBP, NBR) and the Alcohol (Elsevier) co-authored paper - [ ] Obtain and attach the arXiv endorsement from Samuel Gershman (Harvard) as evidence of research community recognition - [ ] Prepare a current CV with ORCID (0009-0001-9272-6735), GitHub (github.com/AmunRaPtah), and personal site (zyco.org) - [ ] Prepare a detailed budget for the $52,500 grant covering living expenses, travel to 2+ AI safety workshops, and computational resources for LLM interpretability experiments - [ ] Draft a one-page project timeline for the 12-month grant period with milestones for each phase - [ ] Identify and list 2-3 specific AI safety workshops or conferences to attend during the grant period - [ ] Prepare contact information for 2-3 references (Kent Berridge, Samuel Gershman, Nathaniel Daw, or Marcelo Mattar) - [ ] Confirm whether the grant requires a letter of recommendation or reference contact details - [ ] Verify whether the grant requires a formal research proposal or if the essays submitted here are sufficient - [ ] Check whether Open Philanthropy requires applicants to be affiliated with a host institution or if independent researchers are eligible - [ ] Confirm the grant's policy on funding researchers who are concurrently enrolled in a degree program EDITOR NOTES - Eligibility risk: The grant is for career transitions into AI safety, and the applicant is concurrently enrolled in an M.Sc. program. Verify whether Open Philanthropy allows grant funding for researchers who are also degree-seeking students, and whether the M.Sc. enrollment strengthens or weakens the case for a career transition grant. - Facts to verify: The applicant's employment timeline shows Synthcare starting March 2026, which is in the future relative to the current date. Confirm the actual current date and adjust employment history accordingly. Also verify the enrollment status at HPI/Potsdam (Winter Semester 2026/27) and whether the grant period would overlap with the first year of the M.Sc. program. - Gap to fill: The application does not specify which AI safety workshops or conferences the applicant would attend. The applicant should identify 2-3 specific events (e.g., the Alignment Workshop, the Interpretability Workshop at NeurIPS, or similar) and name them in the career transition plan. - Gap to fill: The application does not specify which AI safety organizations or PhD programs the applicant would target in phase three of the transition plan. The applicant should name 2-3 specific organizations or programs to demonstrate a concrete job exploration plan. - Framing note: The neurocascade project is the strongest match to the grant's AI safety mission, but its circuit-layer parameters are explicitly labeled illustrative pending real behavioral-data fits. The application should be careful not to overclaim the validation status of neurocascade while still presenting it as evidence of multi-scale modeling capability.