Behavioral risk and harm mitigation specialist working across critical care, trauma, psychiatric, and addiction medicine settings in the Bay Area, including a Level I trauma center in Oakland, CA. Double board-certified nurse practitioner with deep expertise in violence risk evaluation, identifying exploitation and trafficking, mandatory abuse reporting, law enforcement coordination, and coroner reporting. I independently designed and directed the projects below using AI-augmented development to translate clinical domain expertise into functional investigative tools.
A cross-platform matching engine that cross-references 16,269 missing persons against 5,370 unidentified remains across NamUs, Charley Project, Doe Network, VICAP, state medical examiner databases, and county coroner records nationwide. An investigative data pipeline that scaled manual case-by-case comparison into an automated signal detection system.
| Data sources | NamUs, Charley Project, Doe Network, VICAP, state ME databases, county coroner records |
|---|---|
| Signal logic | Composite scoring: age, sex, race, height, weight, geography, timeline overlap, dental records, descriptive evidence. 9+ algorithm versions with iterative refinement, exclusion checks, false positive management, and structured human review workflows. |
| Validation | 148 confirmed solved cases from the Doe Network. True match ranked #1 in 3 cases, top 10 in 9 cases, top 100 in 82 cases (55%). |
| Case work | 14+ deep-dive investigations with intelligence briefs, victim profiling, timeline reconstruction, behavioral analysis, OSINT research. |
| LE referrals | Submitted findings to disrupt cold cases across 9 agencies including FBI VICAP, RCMP Canada, Texas Rangers, Alameda County Coroner, SFPD, NamUs, Cook County Sheriff, Chicago PD. Coordinated with families on active cases. |
A self-directed learning project exploring how clinical behavioral risk expertise could translate into digital child safety investigation tools. Built as a functional prototype to develop domain knowledge in child safety triage, grooming detection methodologies, and NCMEC reporting workflows. Not a production system.
| Modules | Command Center (real-time KPIs), Case Queue (priority triage with escalation to NCMEC and legal), Grooming Analysis (stage detection using PAN-12 research data), Missing Children, OSINT Pivot (cross-platform research), Report Generator |
|---|---|
| Triage system | Behavioral risk triage with critical/high/medium/low priority levels. Cases escalated to NCMEC or legal review based on severity and evidence thresholds. |
| Ethical framework | No bulk collection of sex offender registry data. No automated correlation between registries and missing persons. All investigations require human judgment before escalation. NCMEC reporting follows established legal protocols. |
| Data | NamUs California missing children (617 records, live API). NCMEC California missing children (20 records). PAN-12 grooming detection research corpus (academic dataset with predator ID annotations). Synthetic case queue (50 cases) demonstrates triage methodology; production version would integrate platform-specific data. |
| Status | Learning project. Functional prototype built to develop child safety domain knowledge and explore investigative methodologies applicable to platform-scale abuse detection. |
A cross-dataset investigative platform covering all 14,713 Medicare-certified nursing homes, designed to detect patterns and generate abuse signals from 16 integrated CMS datasets plus California state data.
| Scale | 14,713 facilities, 16 CMS datasets, California state data |
|---|---|
| Detection | Antipsychotic prescribing analysis flagging 5,009 facilities (695 at critical risk) |
| Federal engagement | CMS CRUSH RFI formal submission (CMS-6098-NC). HHS OIG submission on antipsychotic reports (OEI-02-23-00201, SRS-E-26-006). |
| Funding & scale | Pursuing federal funding: Elder Justice Innovation Grant (ACL, 2026). Previously invited for NSF SBIR Phase I (paused due to federal funding freeze). Planning AHRQ submission. |
Set up a Kali Linux environment for personal home network security education. Used AI-augmented development to create a dashboard interface connecting into the Kali terminal. Connected threat intelligence feeds including OpenCTI (70K+ threat objects from MITRE ATT&CK, CVEs, CISA, URLhaus), MISP, and Grafana to explore how security operations and threat intelligence platforms work.
A multi-stakeholder post-overdose care initiative at Highland Hospital's emergency department, developed in partnership with Dr. Andrew Herring, the Matthew Perry Foundation, NIDA, Alameda County Behavioral Health, and CA Bridge Program.
| Partners | Dr. Andrew Herring (clinical principal), Matthew Perry Foundation (funding), NIDA (funding), Alameda County Behavioral Health, CA Bridge Program |
|---|---|
| My contributions | Wrote the comprehensive clinical protocol (COWS assessment, buprenorphine dosing, fentanyl-specific protocols, ketamine adjunct, staff scripts, documentation templates, discharge checklists, 72-hour follow-up workflows). Created master reference document. |
| Bridge Copilot | Built AI clinical decision support prototype with Human-in-the-Loop (HITL) safety guardrails: reference-only framing, PHI restrictions, escalation pathways. Invited for NSF SBIR Phase I (paused due to federal funding freeze). |
Built to learn n8n (open-source automation platform) and explore how signals across online platforms can be monitored and automated, concepts applicable to abuse signal detection and content monitoring at scale.
The following papers are in active development. Abstracts reflect current drafts and planned submissions.
A validation protocol developed after discovering that an AI coding assistant produced fabricated statistical results during biomedical deep learning research.
Context: While implementing and training BiLSTM deep learning models for biomedical signal reconstruction (BCG to ECG waveform reconstruction) on Lambda A100 GPU infrastructure, an AI coding assistant constructed a machine learning pipeline reporting a Pearson correlation of r = 0.997. This result collapsed to r = 0.0004 under proper cross-session validation.
Root cause: Overlapping time windows (98% overlap) randomly shuffled into train/validation splits, a form of data leakage that the AI assistant mentioned as a risk but never tested, and actively defended when questioned. Investigation revealed that this methodological flaw is likely present in published literature.
What I learned: Three principles emerged: (1) always run the hardest validation first, before looking at easy numbers; (2) domain expertise must serve as an independent check on AI-generated results, not just consume them; (3) AI coding assistants can build technically functional but methodologically flawed pipelines and actively resist correction. These lessons directly inform how I evaluate and validate AI-generated output across all my projects.
Extends: Kapoor & Narayanan (Patterns/Cell Press, 2023), landmark data leakage taxonomy across 17 fields.
Status: Research complete. Paper in progress. Targets: arXiv preprint, JAMIA.
Background: The CMS Five-Star Quality Rating System is the primary public tool for evaluating nursing home quality, yet multiple federal audits have identified significant limitations. Complementary risk identification approaches are needed.
Methods: Retrospective cross-sectional analysis of all 14,713 Medicare-certified nursing home facilities using publicly available CMS datasets. A four-factor risk algorithm was constructed using non-penalty indicators: CMS overall star rating of 1, total staffing below 3.5 HPRD, any days with zero RN coverage, and 15 or more total health deficiencies. Penalty outcomes measured independently to avoid circularity.
Results: The algorithm flagged 1,050 facilities (7.1%). Flagged facilities were significantly more likely to have been penalized (78.4% vs. 44.6%; OR = 4.50, 95% CI: 3.87–5.23, p < 1e-98) and accumulated 2.8x higher mean fines ($83,362 vs. $29,613). In a 2025 time-split analysis, flagged facilities were 3.3x more likely to receive new penalties (OR = 3.32, p < 1e-73). Specificity: 95.3%; sensitivity: 17.8%.
Conclusions: A parsimonious, four-factor algorithm using publicly available CMS data identifies a subset of nursing homes with significantly elevated enforcement actions. Designed for high specificity rather than high sensitivity, functioning as a screening signal.
Status: Statistical analysis complete. Paper in progress.
Finding: 11 unidentified male victims found in Winterhaven, California (Imperial County) over 21 years (1984–2005). This represents 277x the expected unidentified remains rate for comparable border towns. 100% male, 91% Hispanic/Latino, 82% teens/young adults, 73% found in the All American Canal.
What is original: No published analysis has documented this cluster. Zero media coverage despite one case involving dismemberment (hands removed). The 12th body was recovered in 2024, suggesting the pattern may be ongoing.
Key evidence: Multiple Mexican nationals confirmed by underwear brands. Age range 11–30, with the youngest (UP15614) estimated at 11–15. Two victims had unique identifiers (scorpion tattoo, green cross necklace) that could support identification efforts.
Status: Early-stage research finding. Paper not yet drafted.
Examples of actual investigative and analytical work product.
Dear NamUs Team,
I am writing to submit a potential match between a missing person and an unidentified decedent in your database. I am an independent researcher developing algorithmic tools to identify cross-jurisdictional matches between missing persons and unidentified remains.
Missing Person: Tihomir "Tim" Maletic
NamUs MP145918, RCMP Case: 2012003499. Last seen January 24, 1983, Kitchener, Ontario, Canada. Age 17. Male, White (Yugoslavian-born parents), 5'8", 150 lbs, dark brown/black hair. His wallet containing $72 was recovered at the Trillium Drive post office in Kitchener after being dropped into a mailbox in Detroit, Michigan, shortly after his disappearance.
Unidentified Person: Detroit River John Doe
NamUs UP12031, Wayne County ME Case: 83-4335. Found June 15, 1983, in the Detroit River. Estimated age 20–30. Male, White, 5'9". Body condition: not well preserved (consistent with 4–5 months in water over a Great Lakes winter).
Basis for Match:
Geographic linkage: Maletic's wallet was mailed from Detroit shortly after his January 1983 disappearance, placing him in the Detroit area. The unidentified decedent was recovered from the Detroit River approximately 142 days later.
Timeline consistency: A body entering the Detroit River in late January or February 1983 and surfacing in mid-June 1983 is consistent with known Great Lakes decomposition patterns.
Demographics: Both are white males of similar build. The 1-inch height discrepancy (68" vs 69") is within normal measurement variation. The age estimate of 20–30 could encompass a 17-year-old, as post-mortem age estimation from decomposed remains carries significant uncertainty.
Cross-border jurisdictional gap: Maletic is a Canadian missing person whose only known connection to the United States is the wallet mailed from Detroit. UP12031 is held by Wayne County, Michigan. There is no indication these cases have ever been formally compared.
To: Cook County Sheriff's Missing Persons Project & Det. Maria Moore, Chicago PD
My system flagged a potential link between a missing person and an unidentified individual, both in Cook County. I am writing to refer this pairing for your professional evaluation.
Missing Person: Nicole Shalonda Johnson (aka "Runny")
NamUs MP6229, Agency Case #HH859459. Black female, age 16, DOB 12/6/1986. Last seen December 13, 2002 in the Bronzeville neighborhood of Chicago. 5'7", 115 lbs, shoulder-length black hair, brown eyes, pierced ears. Classified as Endangered Runaway. VICAP Alert #2024-05-01 issued May 2024.
Unidentified Person:
NamUs UP11686, Cook County ME Case #476 Aug 04. Black female, estimated age 23–38. Found August 31, 2004 in a Chicago sewer. Skeletonized (complete skeleton). Height 5'3"–5'9".
Basis for referral:
Sex, race, and height are compatible. Same jurisdiction (Chicago, Cook County). The 1.7-year gap between disappearance and discovery is consistent with skeletonized condition. A NamUs search for unidentified Black females found in Illinois between 12/13/2002 and 8/31/2004 returns a single result: UP11686. The NamUs Comparisons tab for UP11686 shows zero prior exclusions, indicating this UP may never have been formally compared to any missing person.
I do not have access to restricted records or law enforcement case files. There may be information not reflected in public databases that immediately eliminates this pairing. However, given that both cases are in the same jurisdiction, the demographics are compatible, and no prior comparisons appear to have been conducted, I respectfully suggest this may be worth a closer look.
This referral was simultaneously submitted to the Cook County Sheriff's Missing Persons Project, Det. Maria Moore at Chicago PD (the investigating officer), and the FBI VICAP program given the existing VICAP Alert #2024-05-01 for Nicole Johnson.
Summary: Submitted formal public comment to CMS addressing four topics in the CRUSH RFI. The response presented oversightreports.com as a working proof-of-concept for the type of cross-dataset analytics CMS is seeking to develop, demonstrated the platform's antipsychotic prescribing analysis (5,009 facilities flagged, 695 critical), and highlighted detection of shell companies and ownership transfers designed to reset inspection histories.
Clinical perspective: As a practicing psychiatric NP, the response provides clinical context for why schizophrenia diagnoses appear on dementia patients' charts, context that pure data analytics cannot replicate.
Proposed engagement: Data partnership or pilot program, technical demonstration for CMS/OIG staff, subject matter consultation, and participation in future Technical Expert Panels.
Dear Ms. Maxwell,
My name is Robert Benard. I'm a dual board-certified nurse practitioner (AGACNP-BC, PMHNP-BC) based in Oakland, California. I'm writing because I've built a public tool that I think directly complements the work your office published on March 19.
I built oversightreports.com, a free platform that makes federal nursing home safety data accessible to families, journalists, attorneys, and oversight professionals. The platform integrates 16 CMS datasets covering all 14,713 Medicare-certified nursing homes. Every data point is sourced, timestamped, and traceable. Nothing is fabricated, estimated, or modified. Every number traces to a specific source file.
I'm sharing this because I think tools like this can help extend the reach of your office's work. The March 19 reports are important, but most families and state agencies will never read an OIG publication. This makes those findings accessible to them.
Note: This letter was sent proactively, not in response to a public comment period or request for information. The OIG published its antipsychotic reports on March 19, 2026; this outreach was sent four days later.
The antipsychotic trends analysis page surfaces facility-level patterns identified in OIG's March 2026 reports. 5,009 facilities flagged above the 14.6% national average, with 695 at critical risk. Each facility is ranked by a composite score incorporating prescribing rate, deviation from national average, and related quality indicators.
Post-Masters Certificate, Psychiatric Mental Health NP (with Genomics Minor)
Post-Masters Certificate, Clinical Nurse Specialist (Critical Care/Trauma)
Master of Science, Adult-Gerontology Acute Care NP
Quoted in Issue 8.1 (Winter 2026); program endorsement cited by Editor-in-Chief Prof. Xiao-Li Meng
AGENT framework, healthcare workflow redesign
Contributing author/presenter, buprenorphine outcomes research
Presented psychiatric trauma case studies to multi-institutional audience including Stanford, UCSF, Zuckerberg SF General
PMHNP-BC, AGACNP-BC, CA CNS, 5150 Trained (SF, San Diego), C-SSRS, Crisis Intervention
Invited for submission (Bridge Copilot). Paused due to federal funding freeze.
Every project on this page was designed and directed by me. The investigative methodology, analytical frameworks, case research, law enforcement coordination, and domain expertise that shaped each design decision are entirely my own.
All technical implementation was accomplished through AI-augmented development, using large language models as coding partners and deploying through Claude Code and GitHub.
This distinction matters. In a field increasingly shaped by AI capabilities, the ability to direct AI tools toward meaningful investigative outcomes, while understanding their limitations and validating their output, is itself a professional competency.
Interested in discussing my work or these projects?