MOTIVATION LETTER The fragmented mental health data problem is an architecture problem, not a data problem. A person's self-understanding is scattered across therapy notes, wearable outputs, journal entries, and clinical assessments, each stored in a separate silo with no shared vocabulary. When these streams disagree, the disagreement is treated as noise and averaged away. The psyche-twin project treats disagreement as signal and makes it an explicit edge in a typed knowledge graph. That architecture is the core of the startup I am bringing to the Roseville AI Startup Accelerator 2026. I am Eniola Ayodele Olutogun, a licensed pharmacist and computational researcher with a B.Pharm from the University of Ibadan and current enrollment in the M.Sc. Digital Health program at Hasso Plattner Institute, University of Potsdam. I have built four independent DuckDB-based ingest-to-analyze pipelines across life sciences and social science domains, self-hosted local LLM serving with llama.cpp, and production systems with automated backup and disaster recovery. psyche-twin is a specified architecture, not a slide deck: a typed, multi-scale knowledge graph, an append-only event log, multiple independent evidence streams, and a conversational interface grounded in the graph's current state. I can demo a working prototype within the 8-week window. The problem is specific. A clinician reviewing a patient's history loses time reconciling contradictory self-reports. A patient loses reliability when their own record contradicts what they remember. A wellness user loses information when their mood tracking app cannot talk to their sleep data. psyche-twin gives the user a unified, user-controlled self-model where contradictions are preserved and visible, not erased. I have direct access to the user community. I am a PCN-licensed pharmacist with clinical experience at Ramset Pharmacy and current employment as National Product Manager at Synthcare. I can conduct customer interviews with patients, clinicians, and wellness users during the program. I am open to iterating on the idea based on what those interviews surface. The prototype scope is deliberately small: one evidence stream pair, one disagreement type, one conversational interface. That is testable in 8 weeks. The Roseville accelerator's selection criteria emphasize a real situation where someone loses time, money, information, or reliability, plus access to users and willingness to iterate. psyche-twin matches all four. I am bringing a product architecture with a named user problem, a feasible prototype path, and a community to test it against, not a research project. RESEARCH STATEMENT The psyche-twin project emerged from a specific technical failure I encountered while building computational models of brain circuits. In the neurocascade project, I coupled pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics to behavioral readouts, with Bayesian calibration via PyMC. The system passed 62 of 62 tests. But the data feeding it came from disparate sources with incompatible formats and unstated assumptions. The integration problem was harder than the modeling problem. That experience redirected my attention from modeling the brain to modeling the person. psyche-twin is a typed, multi-scale knowledge-graph architecture for self-modeling. The design is explicit. Multiple independent evidence streams, LLM-derived, assessment-derived, behavioral, and document-derived, fuse into one append-only event log. Each stream keeps its own schema. Disagreement between streams becomes 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. The append-only log ensures the system can always reconstruct what it knew and when it knew it. The technical foundation is already built. I have four independent DuckDB-based ingest-to-analyze corpus and RAG pipelines running across life-sciences, tech/AI/security, and social-science domains. I self-host local LLM serving with llama.cpp and on-demand model swapping. I run production Linux VPS systems with systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery. The graph layer is specified in a typed schema. The conversational interface is a known engineering problem given the graph backend. The validation status is honest. psyche-twin is pre-product, pre-revenue, and pre-user. What exists is a fully specified architecture, a working data-ingestion stack, and a clear prototype path. The first prototype will pair two evidence streams, an LLM-derived stream from conversational input and an assessment-derived stream from a standardized questionnaire, and surface disagreements between them as explicit graph edges. The conversational interface will answer questions grounded in the current graph state. That is the demo. The scientific motivation is grounded in my prior work. My CCT model, a tripartite pharmacological framework for reward-memory encoding prevention in addiction, required integrating dopaminergic RPE signals, NMDAR-dependent LTP, and affective contrast into a single ODE system with Bayesian MCMC calibration. All five pre-registered hypotheses were confirmed. The integration challenge in that project taught me that the data architecture determines what the model can see. psyche-twin applies that lesson to the individual user. The market need is documented. Mental health data is fragmented across providers, apps, and self-reports. Users lose time reconciling contradictions. Clinicians lose reliability when records disagree. The current market response is to build more silos. psyche-twin builds a unified layer where the user owns the graph and the contradictions are preserved. The 8-week accelerator timeline is feasible. Week 1-2: finalize the typed schema for two evidence streams. Week 3-4: build the ingest pipeline for conversational and assessment data. Week 5-6: implement the disagreement-edge logic and the conversational interface. Week 7-8: conduct user interviews with my clinical network and iterate on the prototype. The scope is deliberately small. That is the point. SHORT ANSWER ESSAY: USER ACCESS My user access is direct and professional. I am a PCN-licensed pharmacist with clinical experience at Ramset Pharmacy from January to March 2026, and I currently serve as National Product Manager at Synthcare, a role I began in March 2026. My professional network includes practicing clinicians, hospital pharmacists, and patients in Nigeria. During the accelerator, I will conduct structured customer interviews with at least three groups: clinicians who review patient histories, patients who track their own mental health data, and wellness users who use self-tracking apps. Each group has a distinct pain point. Clinicians lose time reconciling contradictory self-reports. Patients lose reliability when their records disagree with their memory. Wellness users lose information when their apps cannot share data. I can recruit interviewees from my existing professional network within the first week of the program. I have no regulatory or confidentiality barriers to these interviews because I will interview in my capacity as an independent researcher, not as a clinician accessing patient records. SHORT ANSWER ESSAY: ITERATION PLAN My iteration plan is structured around the accelerator's 8-week timeline. In weeks 1 and 2, I will conduct 10 to 15 customer interviews using the problem-interview format, asking about specific instances where fragmented mental health data caused lost time, lost money, lost information, or lost reliability. In week 3, I will map the interview results against the psyche-twin architecture and identify the single most common pain point. In weeks 4 and 5, I will build the smallest prototype that addresses that pain point, pairing one conversational evidence stream with one assessment stream and surfacing disagreements as graph edges. In week 6, I will return to the same interviewees with the prototype and record their reactions. In weeks 7 and 8, I will iterate on the prototype based on that feedback and prepare the demo. My prior work demonstrates this discipline. In the ergofluids project, I pre-registered a gated validation pipeline. The first real-data gate did not meet its primary criterion, and I reported that result directly rather than reframing it. I will apply the same standard to customer feedback. If the interviews show the problem is not where I think it is, I will change the prototype. CHECKLIST - [ ] Confirm Roseville AI Startup Accelerator 2026 application deadline from the programme website - [ ] Verify whether the accelerator requires a company registration or accepts pre-incorporation applicants - [ ] Prepare a 2-minute demo video of the psyche-twin prototype concept, showing the typed knowledge graph and conversational interface - [ ] Prepare a one-page technical architecture document for psyche-twin, including the typed schema and evidence-stream specifications - [ ] Prepare a customer interview guide for the three user groups: clinicians, patients, wellness users - [ ] Prepare a list of 15 potential interviewees from the Synthcare and Ramset Pharmacy professional networks - [ ] Verify the M.Sc. Digital Health enrollment status at HPI/Potsdam for the Winter Semester 2026/27 and include proof of enrollment - [ ] Confirm the ORCID iD (0009-0001-9272-6735) and GitHub URL are listed on the application - [ ] Draft a 30-second elevator pitch for the psyche-twin product, not the research project - [ ] Prepare a slide deck of no more than 10 slides covering problem, prototype, user access, and iteration plan EDITOR NOTES - Eligibility risk: the accelerator may require a registered company or a founding team; the profile shows an independent researcher with no incorporation. Verify before submitting. - The psyche-twin project is pre-product and pre-revenue. Do not claim validated IP, product-market fit, or revenue anywhere in the application. The profile explicitly states this is a venture behind a real-data validation gate. - The applicant's current employment as National Product Manager at Synthcare may present a conflict of interest if the accelerator expects full-time commitment. Confirm the time commitment required and whether the role can be paused or reduced during the 8-week program. - The M.Sc. Digital Health enrollment at HPI/Potsdam is listed for Winter Semester 2026/27. If the accelerator runs in August 2026, the applicant will be pre-enrollment. Confirm whether the program requires full-time availability during that period. - The applicant's clinical network is in Nigeria. If the accelerator expects US or European user interviews, the applicant must either travel or conduct remote interviews. Confirm the accelerator's geographic expectations for customer discovery. - The profile lists the applicant as 29 years old and Nigerian. Some accelerators have age or geographic quotas. Confirm the Roseville accelerator has no such restrictions.