Draft — Sacramento AI-Native Startup Accelerator

MOTIVATION LETTER

The everyday problem is that people cannot see themselves clearly. Mental health care depends on self-report, and self-report is filtered through memory, mood, and social desirability. The psyche-twin project builds a typed, multi-scale knowledge graph that fuses multiple independent evidence streams, an LLM-derived stream, an assessment-derived stream, a behavioral stream, and a document-derived stream, into one append-only event log. When streams disagree, the disagreement becomes an explicit graph edge rather than being averaged away. A first-person conversational interface sits on top of the graph and generates responses grounded in the graph's current state. This is AI-native by construction: the product is the graph, and the graph is the model of the user.

The Sacramento AI-Native Startup Accelerator asks for everyday problem solvers who put AI at the core of their product. psyche-twin qualifies on both counts. The problem is universal: self-understanding is a prerequisite for mental health, and current tools do not provide it. The AI integration is not a feature layer; the entire architecture is an AI system. The multi-scale knowledge graph is built with typed edges, append-only logging, and explicit disagreement representation, all of which are AI-infrastructure decisions, not product decisions bolted on later.

My technical foundation is documented across multiple independent research lines. I built the CCT model, a tripartite pharmacological framework for reward-memory encoding prevention in addiction, with a coupled three-axis ODE model calibrated via Bayesian MCMC using PyMC DEMetropolisZ across 14 free parameters. All five pre-registered hypotheses were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. I built TOPOLOGIX, an ESM-2 protein-language-model delta-embedding pipeline with Morgan fingerprints and a Random Forest classifier, achieving AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, beating structure-based baselines like mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus approximately 18 percent for structure-limited tools. I built neurocascade, a receptor-to-behavior brain-circuit simulation engine coupling pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics, with 62 of 62 tests passing. I also ran a pre-registered replication study on hERG cardiotoxicity topology that found topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782, a negative result I reported directly rather than reframing.

The accelerator's selection criteria include problem clarity, founder potential, AI integration, feasibility, impact, and fit with the everyday problem solvers ethos. psyche-twin addresses problem clarity directly: the problem is self-knowledge, and the solution is a graph that makes the user's own evidence streams legible. Founder potential is demonstrated by my track record of independent, pre-registered, falsifiable research across neuroscience, protein ML, and dynamical systems. Feasibility is grounded in my applied infrastructure work: four independent DuckDB-based ingest-to-analyze pipelines, self-hosted local LLM serving with llama.cpp, and production systems operations including Linux VPS, systemd, Caddy TLS, and automated backup and disaster recovery. Impact is measurable: the graph architecture is designed to surface contradictions in self-modeling that standard assessment tools hide.

The accelerator's clean-tech focus is not a barrier. The AI training and AI-native startup methodology transfer directly to a mental-health product built on knowledge-graph infrastructure. The venture is at the stage where accelerator guidance on market entry and scalability would have the highest use. I am enrolled in the M.Sc. Digital Health program at Hasso Plattner Institute and University of Potsdam starting Winter Semester 2026/27, which provides academic grounding in digital health systems while the venture develops independently.

psyche-twin is the right project for this accelerator because it is an everyday problem solver with AI at its core, and it is at the stage where structured startup training matters most.

RESEARCH STATEMENT

The psyche-twin project is a typed, multi-scale knowledge-graph architecture for self-modeling. The core research question is whether a system can represent a person's internal state more faithfully by fusing multiple independent evidence streams into a single append-only event log, where disagreement between streams is preserved as an explicit graph edge rather than averaged away. This is a computational question about representation, not a product question. The answer determines whether self-modeling systems can be built on honest disagreement rather than on smoothed consensus.

The architecture has four evidence streams. The LLM-derived stream extracts structured claims from conversational input. The assessment-derived stream ingests standardized instrument scores. The behavioral stream logs actions and their contexts. The document-derived stream ingests external records. Each stream writes typed events to an append-only log. The graph layer builds typed edges between events, entities, and claims. When two streams produce conflicting evidence about the same entity, the conflict is stored as a distinct edge type with both sides preserved. The conversational interface reads the graph state and generates responses grounded in that state, including explicit acknowledgment of unresolved conflicts.

The technical foundation for this work comes from my prior research. The CCT model demonstrated that a coupled ODE system with Bayesian calibration can produce falsifiable predictions about reward-memory encoding, with all five pre-registered hypotheses confirmed. The hERG topology study demonstrated the opposite: a pre-registered, powered replication found that bipartite persistent homology does not beat a plain descriptor baseline for cardiotoxicity prediction, AUROC 0.8426 versus 0.8782. That negative result is as important as the positive ones. It established that I report what the data says, not what I hoped it would say. The TOPOLOGIX project then pivoted on that negative result, moving from interface geometry to sequence representation and achieving AUROC 0.804 plus or minus 0.025 on the Platinum benchmark. The pattern across these projects is consistent: pre-register the hypothesis, run the analysis, report the outcome, and adjust the next hypothesis accordingly.

psyche-twin applies the same discipline to self-modeling. The pre-registration question is whether disagreement-preserving graph representations produce more accurate self-models than consensus-based representations. The evaluation metric is whether the graph state predicts subsequent behavioral outcomes better than a baseline that averages across streams. The system is currently at the architecture and implementation stage. The graph schema is typed, the event log is append-only, and the conversational interface is functional. What remains is the validation study: a longitudinal dataset where multiple evidence streams are collected over time, and the disagreement-preserving graph is tested against a consensus baseline for predictive accuracy.

The accelerator's AI-native criterion is satisfied by the architecture itself. The graph is not a database behind an AI product; the graph is the AI system. The conversational interface is a read layer over the graph state. The LLM-derived stream is one of four evidence sources, not the sole source of truth. This design prevents the common failure mode where an LLM generates plausible but ungrounded self-descriptions. Every claim in the graph is traceable to a typed event in the append-only log.

Feasibility is grounded in my applied infrastructure experience. I have built four independent DuckDB-based ingest-to-analyze pipelines across life sciences, tech and AI security, and social science domains. I self-host local LLM serving with llama.cpp and on-demand model swapping. I run production systems on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery. The psyche-twin stack uses Supabase and Postgres for the graph store, which I have deployed in prior work.

The impact case is direct. Mental health assessment today relies on instruments that collapse multidimensional states into scalar scores. Those scores hide the contradictions that matter clinically. A person can report low depression on a scale while their behavioral stream shows social withdrawal and their conversational stream shows rumination. A consensus-based system averages these into a moderate score. A disagreement-preserving graph keeps all three signals visible and makes the contradiction itself a data point. That is the research contribution, and it is the product differentiator.

The accelerator's training focus on AI-native startup methodology is the right next step. The research is at the stage where market entry questions, user acquisition for a self-modeling tool, and scalability of the graph infrastructure become as important as the representation question. The accelerator's emphasis on everyday problem solvers matches the product's target user: anyone who wants to see themselves more clearly, not just clinical populations.

SHORT ANSWER ESSAYS

Problem clarity: The everyday problem is that self-knowledge is unreliable because it depends on a single, filtered evidence stream. People answer questionnaires, talk to therapists, and reflect on their behavior, but these streams are never integrated. Contradictions between what someone says, what they score, and what they do are hidden. psyche-twin solves this by building a knowledge graph that fuses multiple evidence streams into one append-only event log, preserving disagreements as explicit graph edges. The conversational interface reads the graph state and surfaces those contradictions. The problem is universal, the solution is concrete, and the AI is the mechanism, not an add-on.

Founder potential: My research record demonstrates sustained independent execution. I built the CCT model with Bayesian MCMC calibration across 14 free parameters and confirmed all five pre-registered hypotheses. I ran a pre-registered replication on hERG cardiotoxicity that produced a negative result, AUROC 0.8426 versus 0.8782, and I reported it directly. I built TOPOLOGIX on that negative result and achieved AUROC 0.804 on the Platinum benchmark. I built neurocascade with 62 of 62 tests passing. I am coachable because I follow pre-registered protocols and adjust hypotheses when data contradicts them. I am committed because I have pursued this work independently without institutional funding.

AI integration: AI is the product, not a feature. The knowledge graph is built from LLM-derived, assessment-derived, behavioral, and document-derived evidence streams. The graph itself is the model of the user. The conversational interface generates responses grounded in the graph state. Disagreement between streams is stored as a typed edge, which means the system can reason about contradictions rather than hiding them. The LLM is one of four evidence sources, not the sole source of truth. Every claim in the graph is traceable to a typed event in the append-only log.

Feasibility: The architecture is implemented. The graph schema is typed, the event log is append-only, and the conversational interface is functional. My infrastructure experience includes four DuckDB-based ingest-to-analyze pipelines, self-hosted LLM serving with llama.cpp, and production systems on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backups. The remaining work is a longitudinal validation study comparing disagreement-preserving graphs against consensus baselines for predictive accuracy. The accelerator's training on market entry and scalability would close the gap between research prototype and deployable product.

Impact: The target community is anyone who struggles with self-understanding, which is everyone. The measurable impact is whether the graph state predicts subsequent behavioral outcomes better than consensus-based baselines. If it does, the product changes how mental health assessment works. If it does not, the negative result still contributes to the field by ruling out a plausible architecture. Either outcome is publishable, and either outcome advances the state of self-modeling systems.

Fit with accelerator focus: The accelerator asks for everyday problem solvers with AI-native products. psyche-twin is an everyday problem solver because self-knowledge is a universal need. It is AI-native because the graph is the AI system. The clean-tech focus of the accelerator does not conflict with a mental-health product; the AI training and startup methodology transfer directly. The venture is at the stage where structured guidance on market entry and scalability has the highest use.

CHECKLIST

- [ ] Complete the online application form at https://www.humanbulb.org/rosevilleacceleratorapplication
- [ ] Verify the application deadline on the programme website
- [ ] Confirm eligibility for independent researchers without a registered company entity
- [ ] Prepare a pitch deck summarizing psyche-twin architecture, validation status, and market entry plan
- [ ] Prepare a demo video or screen recording of the conversational interface reading the graph state
- [ ] Gather verification for the M.Sc. Digital Health enrollment at Hasso Plattner Institute, Winter Semester 2026/27
- [ ] Prepare a one-page technical appendix describing the graph schema, event log design, and disagreement edge types
- [ ] List the four evidence streams with concrete examples of each stream's output format
- [ ] Prepare a timeline for the longitudinal validation study, including participant recruitment and IRB considerations
- [ ] Identify the specific accelerator training modules most relevant to psyche-twin and note them in the application
- [ ] Prepare contact information for one reference who can speak to technical execution (e.g., a collaborator or former supervisor)
- [ ] Confirm whether the accelerator requires a working prototype or accepts a research-stage system with functional architecture

EDITOR NOTES

- Eligibility risk: The accelerator may require a registered company or a founding team. Eniola is an independent researcher without a stated corporate entity. Verify whether the accelerator accepts solo founders without incorporation, or whether a Delaware C-corp or Nigerian entity needs to be formed before application.
- Fact verification: The profile states enrollment in M.Sc. Digital Health at HPI/Potsdam for Winter Semester 2026/27. Confirm the exact program name and start date, as the application may ask for proof of enrollment or academic affiliation.
- Gap to fill: The profile does not specify whether psyche-twin has any user testing, pilot users, or letters of intent from potential customers. The accelerator's feasibility and impact criteria may require evidence of demand. Eniola should insert any user feedback, waitlist numbers, or pilot conversations if they exist.
- Gap to fill: The profile does not state whether psyche-twin has been pre-registered as a study. Given Eniola's track record of pre-registration in prior projects, the application should state the pre-registration status for the validation study, or commit to a pre-registration date.
- Fit note: The strategy notes flag that the accelerator focuses on clean tech and that geography is not a priority. The application leans on the AI-native and everyday problem solver criteria. If the accelerator's clean-tech focus is strict, the application may be rejected on sector grounds regardless of quality. Consider whether a backup application to a health-tech or AI-focused accelerator is warranted.

MOTIVATION LETTER

The everyday problem is that people cannot see themselves clearly. Mental health care depends on self-report, and self-report is filtered through memory, mood, and social desirability. The psyche-twin project builds a typed, multi-scale knowledge graph that fuses multiple independent evidence streams, an LLM-derived stream, an assessment-derived stream, a behavioral stream, and a document-derived stream, into one append-only event log. When streams disagree, the disagreement becomes an explicit graph edge rather than being averaged away. A first-person conversational interface sits on top of the graph and generates responses grounded in the graph's current state. This is AI-native by construction: the product is the graph, and the graph is the model of the user.

The Sacramento AI-Native Startup Accelerator asks for everyday problem solvers who put AI at the core of their product. psyche-twin qualifies on both counts. The problem is universal: self-understanding is a prerequisite for mental health, and current tools do not provide it. The AI integration is not a feature layer; the entire architecture is an AI system. The multi-scale knowledge graph is built with typed edges, append-only logging, and explicit disagreement representation, all of which are AI-infrastructure decisions, not product decisions bolted on later.

My technical foundation is documented across multiple independent research lines. I built the CCT model, a tripartite pharmacological framework for reward-memory encoding prevention in addiction, with a coupled three-axis ODE model calibrated via Bayesian MCMC using PyMC DEMetropolisZ across 14 free parameters. All five pre-registered hypotheses were confirmed, with posterior super-additivity of 13 to 22 percentage points across model versions. I built TOPOLOGIX, an ESM-2 protein-language-model delta-embedding pipeline with Morgan fingerprints and a Random Forest classifier, achieving AUROC 0.804 plus or minus 0.025 on the Platinum benchmark of 553 mutations, beating structure-based baselines like mCSM-lig at approximately 0.70 while covering 100 percent of mutations versus approximately 18 percent for structure-limited tools. I built neurocascade, a receptor-to-behavior brain-circuit simulation engine coupling pharmacokinetics to receptor binding to Wilson-Cowan circuit dynamics, with 62 of 62 tests passing. I also ran a pre-registered replication study on hERG cardiotoxicity topology that found topological features do not beat a plain descriptor baseline, AUROC 0.8426 versus 0.8782, a negative result I reported directly rather than reframing.

The accelerator's selection criteria include problem clarity, founder potential, AI integration, feasibility, impact, and fit with the everyday problem solvers ethos. psyche-twin addresses problem clarity directly: the problem is self-knowledge, and the solution is a graph that makes the user's own evidence streams legible. Founder potential is demonstrated by my track record of independent, pre-registered, falsifiable research across neuroscience, protein ML, and dynamical systems. Feasibility is grounded in my applied infrastructure work: four independent DuckDB-based ingest-to-analyze pipelines, self-hosted local LLM serving with llama.cpp, and production systems operations including Linux VPS, systemd, Caddy TLS, and automated backup and disaster recovery. Impact is measurable: the graph architecture is designed to surface contradictions in self-modeling that standard assessment tools hide.

The accelerator's clean-tech focus is not a barrier. The AI training and AI-native startup methodology transfer directly to a mental-health product built on knowledge-graph infrastructure. The venture is at the stage where accelerator guidance on market entry and scalability would have the highest use. I am enrolled in the M.Sc. Digital Health program at Hasso Plattner Institute and University of Potsdam starting Winter Semester 2026/27, which provides academic grounding in digital health systems while the venture develops independently.

psyche-twin is the right project for this accelerator because it is an everyday problem solver with AI at its core, and it is at the stage where structured startup training matters most.

RESEARCH STATEMENT

The psyche-twin project is a typed, multi-scale knowledge-graph architecture for self-modeling. The core research question is whether a system can represent a person's internal state more faithfully by fusing multiple independent evidence streams into a single append-only event log, where disagreement between streams is preserved as an explicit graph edge rather than averaged away. This is a computational question about representation, not a product question. The answer determines whether self-modeling systems can be built on honest disagreement rather than on smoothed consensus.

The architecture has four evidence streams. The LLM-derived stream extracts structured claims from conversational input. The assessment-derived stream ingests standardized instrument scores. The behavioral stream logs actions and their contexts. The document-derived stream ingests external records. Each stream writes typed events to an append-only log. The graph layer builds typed edges between events, entities, and claims. When two streams produce conflicting evidence about the same entity, the conflict is stored as a distinct edge type with both sides preserved. The conversational interface reads the graph state and generates responses grounded in that state, including explicit acknowledgment of unresolved conflicts.

The technical foundation for this work comes from my prior research. The CCT model demonstrated that a coupled ODE system with Bayesian calibration can produce falsifiable predictions about reward-memory encoding, with all five pre-registered hypotheses confirmed. The hERG topology study demonstrated the opposite: a pre-registered, powered replication found that bipartite persistent homology does not beat a plain descriptor baseline for cardiotoxicity prediction, AUROC 0.8426 versus 0.8782. That negative result is as important as the positive ones. It established that I report what the data says, not what I hoped it would say. The TOPOLOGIX project then pivoted on that negative result, moving from interface geometry to sequence representation and achieving AUROC 0.804 plus or minus 0.025 on the Platinum benchmark. The pattern across these projects is consistent: pre-register the hypothesis, run the analysis, report the outcome, and adjust the next hypothesis accordingly.

psyche-twin applies the same discipline to self-modeling. The pre-registration question is whether disagreement-preserving graph representations produce more accurate self-models than consensus-based representations. The evaluation metric is whether the graph state predicts subsequent behavioral outcomes better than a baseline that averages across streams. The system is currently at the architecture and implementation stage. The graph schema is typed, the event log is append-only, and the conversational interface is functional. What remains is the validation study: a longitudinal dataset where multiple evidence streams are collected over time, and the disagreement-preserving graph is tested against a consensus baseline for predictive accuracy.

The accelerator's AI-native criterion is satisfied by the architecture itself. The graph is not a database behind an AI product; the graph is the AI system. The conversational interface is a read layer over the graph state. The LLM-derived stream is one of four evidence sources, not the sole source of truth. This design prevents the common failure mode where an LLM generates plausible but ungrounded self-descriptions. Every claim in the graph is traceable to a typed event in the append-only log.

Feasibility is grounded in my applied infrastructure experience. I have built four independent DuckDB-based ingest-to-analyze pipelines across life sciences, tech and AI security, and social science domains. I self-host local LLM serving with llama.cpp and on-demand model swapping. I run production systems on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backup and disaster recovery. The psyche-twin stack uses Supabase and Postgres for the graph store, which I have deployed in prior work.

The impact case is direct. Mental health assessment today relies on instruments that collapse multidimensional states into scalar scores. Those scores hide the contradictions that matter clinically. A person can report low depression on a scale while their behavioral stream shows social withdrawal and their conversational stream shows rumination. A consensus-based system averages these into a moderate score. A disagreement-preserving graph keeps all three signals visible and makes the contradiction itself a data point. That is the research contribution, and it is the product differentiator.

The accelerator's training focus on AI-native startup methodology is the right next step. The research is at the stage where market entry questions, user acquisition for a self-modeling tool, and scalability of the graph infrastructure become as important as the representation question. The accelerator's emphasis on everyday problem solvers matches the product's target user: anyone who wants to see themselves more clearly, not just clinical populations.

SHORT ANSWER ESSAYS

Problem clarity: The everyday problem is that self-knowledge is unreliable because it depends on a single, filtered evidence stream. People answer questionnaires, talk to therapists, and reflect on their behavior, but these streams are never integrated. Contradictions between what someone says, what they score, and what they do are hidden. psyche-twin solves this by building a knowledge graph that fuses multiple evidence streams into one append-only event log, preserving disagreements as explicit graph edges. The conversational interface reads the graph state and surfaces those contradictions. The problem is universal, the solution is concrete, and the AI is the mechanism, not an add-on.

Founder potential: My research record demonstrates sustained independent execution. I built the CCT model with Bayesian MCMC calibration across 14 free parameters and confirmed all five pre-registered hypotheses. I ran a pre-registered replication on hERG cardiotoxicity that produced a negative result, AUROC 0.8426 versus 0.8782, and I reported it directly. I built TOPOLOGIX on that negative result and achieved AUROC 0.804 on the Platinum benchmark. I built neurocascade with 62 of 62 tests passing. I am coachable because I follow pre-registered protocols and adjust hypotheses when data contradicts them. I am committed because I have pursued this work independently without institutional funding.

AI integration: AI is the product, not a feature. The knowledge graph is built from LLM-derived, assessment-derived, behavioral, and document-derived evidence streams. The graph itself is the model of the user. The conversational interface generates responses grounded in the graph state. Disagreement between streams is stored as a typed edge, which means the system can reason about contradictions rather than hiding them. The LLM is one of four evidence sources, not the sole source of truth. Every claim in the graph is traceable to a typed event in the append-only log.

Feasibility: The architecture is implemented. The graph schema is typed, the event log is append-only, and the conversational interface is functional. My infrastructure experience includes four DuckDB-based ingest-to-analyze pipelines, self-hosted LLM serving with llama.cpp, and production systems on Linux VPS with systemd, Caddy TLS, CI/CD, and automated backups. The remaining work is a longitudinal validation study comparing disagreement-preserving graphs against consensus baselines for predictive accuracy. The accelerator's training on market entry and scalability would close the gap between research prototype and deployable product.

Impact: The target community is anyone who struggles with self-understanding, which is everyone. The measurable impact is whether the graph state predicts subsequent behavioral outcomes better than consensus-based baselines. If it does, the product changes how mental health assessment works. If it does not, the negative result still contributes to the field by ruling out a plausible architecture. Either outcome is publishable, and either outcome advances the state of self-modeling systems.

Fit with accelerator focus: The accelerator asks for everyday problem solvers with AI-native products. psyche-twin is an everyday problem solver because self-knowledge is a universal need. It is AI-native because the graph is the AI system. The clean-tech focus of the accelerator does not conflict with a mental-health product; the AI training and startup methodology transfer directly. The venture is at the stage where structured guidance on market entry and scalability has the highest use.

CHECKLIST

- [ ] Complete the online application form at https://www.humanbulb.org/rosevilleacceleratorapplication
- [ ] Verify the application deadline on the programme website
- [ ] Confirm eligibility for independent researchers without a registered company entity
- [ ] Prepare a pitch deck summarizing psyche-twin architecture, validation status, and market entry plan
- [ ] Prepare a demo video or screen recording of the conversational interface reading the graph state
- [ ] Gather verification for the M.Sc. Digital Health enrollment at Hasso Plattner Institute, Winter Semester 2026/27
- [ ] Prepare a one-page technical appendix describing the graph schema, event log design, and disagreement edge types
- [ ] List the four evidence streams with concrete examples of each stream's output format
- [ ] Prepare a timeline for the longitudinal validation study, including participant recruitment and IRB considerations
- [ ] Identify the specific accelerator training modules most relevant to psyche-twin and note them in the application
- [ ] Prepare contact information for one reference who can speak to technical execution (e.g., a collaborator or former supervisor)
- [ ] Confirm whether the accelerator requires a working prototype or accepts a research-stage system with functional architecture

EDITOR NOTES

- Eligibility risk: The accelerator may require a registered company or a founding team. Eniola is an independent researcher without a stated corporate entity. Verify whether the accelerator accepts solo founders without incorporation, or whether a Delaware C-corp or Nigerian entity needs to be formed before application.
- Fact verification: The profile states enrollment in M.Sc. Digital Health at HPI/Potsdam for Winter Semester 2026/27. Confirm the exact program name and start date, as the application may ask for proof of enrollment or academic affiliation.
- Gap to fill: The profile does not specify whether psyche-twin has any user testing, pilot users, or letters of intent from potential customers. The accelerator's feasibility and impact criteria may require evidence of demand. Eniola should insert any user feedback, waitlist numbers, or pilot conversations if they exist.
- Gap to fill: The profile does not state whether psyche-twin has been pre-registered as a study. Given Eniola's track record of pre-registration in prior projects, the application should state the pre-registration status for the validation study, or commit to a pre-registration date.
- Fit note: The strategy notes flag that the accelerator focuses on clean tech and that geography is not a priority. The application leans on the AI-native and everyday problem solver criteria. If the accelerator's clean-tech focus is strict, the application may be rejected on sector grounds regardless of quality. Consider whether a backup application to a health-tech or AI-focused accelerator is warranted.

AI Draft — Sacramento AI-Native Startup Accelerator