Marine Scientist · Environmental Steward · Computational Biologist
Part of SCMTS's volunteer trail crews building and maintaining trails at Wilder Ranch State Park and Henry Cowell Redwoods State Park, contributing to new public trail access in the Santa Cruz Mountains.
As Trail Committee Chair for Advocates for Nisene Marks State Park, I lead conservation stewardship initiatives combining field leadership with geospatial technology to protect old-growth redwood ecosystems.
Trail stewardship and habitat restoration at Elkhorn Slough National Estuarine Research Reserve, maintaining visitor access through coastal wetland environments while supporting the reserve's ecological mission.
Volunteer with the Stewardship Trails & Access Crew, maintaining roads, trails, and ecological integrity across Big Sur Land Trust's protected coastal and mountain landscapes.
Conservation volunteer on the 20,000-acre Santa Lucia Preserve in Carmel, working alongside the Conservancy to maintain trails and support science-guided land management on permanently protected wildlands.
What sharks do when no one is watching
Tracking animals without drift
Telemetry where there is none
Speeding up deep-sea data ingestion
Turning “zombie” urchins into a fishery
Which growth curves can be trusted
Mark-recapture without tagging
Surveying an unfamiliar coast
Getting to where the data is
The sharks in this water are 3D models I built for my own projects, rigged and animated here rather than dropped in from a library.
Ocean Predator Ecology Lab, CSUMB
Spring 2023 – Present · Monterey Bay, California
An archive running 2012–2026, a population of individually recognizable white sharks, and one question underneath all of it: how much of an animal’s health can you read off its body? This work turns that footage into measurements, and runs the fieldwork that keeps producing more of it.
NEPSS: White Shark Morphometric AnalysisMarch 2026
Implementations
Monterey Bay Aquarium Research Institute
June – August 2026 · Moss Landing, California
FathomNet is the open image database the deep-sea machine learning community trains on. When I arrived it could only store bounding boxes, so a decade of ROV imagery could say where an animal was but never what shape it was. Over one summer I extended it end to end to hold points and segmentation masks, then filled the existing corpus, scaling from a SAM 2 pilot to a full-corpus SAM 3 run. Every mask is graded rather than accepted, with rejects held back under an explicit deferred count so the accept rate cannot be quietly inflated.
Recording of the final presentation being added
Paper publishing to MBARI’s intern papers site, September 2026
The schema and the orchestrator
Jorgensen Lab, CSUMB
May – August 2026 · CSUMB
A biologging tag reconstructs an animal’s track by dead reckoning from accelerometer and magnetometer data. Over hours that track drifts badly, and a behavior classifier trained on it can quietly learn to recognize the individual animal instead of the behavior. anchor is the correction for both problems.
Interchange formats and the smoother
Independent, with Jorgensen Lab and partner catalogues
2025 – Present · Monterey Bay
A harbor porpoise surfaces for about a second. A sevengill shark comes over the side of a boat once and goes back. In both cases a photograph is all you get, and the question is the same one: have we seen this animal before? Answer it well enough and a sighting becomes a resighting, which is what turns photographs into a mark-recapture study without ever putting a tag on anything. The machinery is shared. What changes between species is which part of the animal carries the signature.
The pipeline and the catalogue
Aquaculture Lab, Moss Landing Marine Laboratories
Spring 2024 – Present · Moss Landing, California
Purple urchins strip a kelp forest and then sit in the barrens they made, starving, with gonads too small to sell. If you can fatten those animals in tanks, removing them stops being a cost centre and starts being a fishery. Two experiments test whether that works, run end to end from hypothesis through manuscript. Related work looks at eelgrass (Zostera marina) as a cornerstone for California aquaculture and blue carbon.
How many urchins can share a tank before gonad development and survival suffer? The answer sets the economics of the whole idea.
Whether a short finishing diet can push roe from starved to market grade, assessed by Texture Profile Analysis alongside carotenoid application and pellet processing.
Strongylocentrotus purpuratus, collected from barrens where they have grazed the kelp down and stalled at low gonad index.
Aquaculture America: Carotenoid Application & Pellet ProcessingFebruary 2026
Jue Lab, CSUMB
Spring 2025 – Present · CSUMB
92 bacterial strains across 6 experimental groups run by 3 operators, and one question: will any of them eat a pesticide. The real question underneath is which strains metabolize the compound rather than merely tolerating it, and how confident that call can be given how noisy plate reader data is. Growth curves come off the reader by the thousand, so the analysis had to be automated before the biology could be answered at all.
Implementation
Galápagos & California
2024 – Present
A commercial VHF telemetry receiver costs $695, which is the real reason so much of the world has no telemetry coverage. I built one for under $100, deployed it in the Galápagos and across California sites, measured what it actually detects in the field rather than on a datasheet, and taught local teams to build and repair their own.
Antenna geometry and gain measured on site rather than assumed, because a station is only as good as the day it was set up
Build it yourself
South African coastal waters
July – August 2024
Six weeks of working dives on a coastline none of us knew, running whatever survey the day called for. It is the clearest lesson I have had in how much of marine science is method discipline under conditions you did not choose.
California & Baja California, Mexico
2024 – Present
The hours behind every other bubble on this page. Diving and vessel work are not a separate interest; they are what makes the rest of the research possible on the days the weather cooperates.
Designed and conducted “Thermal Thresholds: The Role of Temperature Fluctuations in Shaping Kelp Forest Health at San Carlos Beach, Monterey, CA” as the research component of AAUS certification.
From kelp canopy surveys to offshore tagging operations, every day on the water has reinforced the same lesson: marine field research is where assumptions meet reality. The ocean does not care about your methods section. It demands adaptability, teamwork, and respect for the unpredictable.
These are not ten separate projects. The relay stations and the biologging tags produce the raw records, the annotation platform turns imagery into labels, the pose and re-identification models consume those labels, anchor reconstructs where the animal actually went, and the FathomNet work is the interchange format that lets any of it be read by anyone else. Every piece exists because the piece upstream of it produced something the piece downstream could not yet use.
The through line is the same in all of it: measure the animal rather than count it, and report the number that survives an honest split.
Real-time HAB monitoring combining NOAA VIIRS satellite data, buoy measurements, and ensemble ML. Built 3 custom REST APIs, dual dashboards for growers and public safety, and automated alert system.
Custom-built edge station: Raspberry Pi 5 + RTL-SDR Blog V4 + SIM7028 NB-IoT HAT. Each unit operates autonomously in the field for weeks, detecting VHF wildlife tags across the 151 MHz band.
927-line multi-channel FFT-based detection engine. Processes 256K IQ samples through adaptive baseline tracking, pattern validation with ghost filtering, and coefficient-of-variation analysis to distinguish true repeating tags (CV < 0.15) from noise.
Dual dashboard architecture: local Flask station dashboard for real-time tag detection and system health, plus a centralized FastAPI server with 35+ REST endpoints, 8-table SQLite schema, HMAC auth, and Leaflet geographic mapping for multi-station fleet monitoring.
Automatic WiFi-to-NB-IoT failover for network resilience. SQLite-backed offline queue buffers detections during connectivity gaps. Batched events use single-letter JSON keys to achieve ~650 bytes per cellular alert. Exponential backoff (2s→60s) with connection state machine tracking wifi/nbiot/offline states.
Field-tested at Monterey Bay coastal sites. Yagi directional antenna achieves 1,700+ ft max range with 62% detection rate. Auto-calibration sweep optimizes SDR gain (15–45 dB) on boot. Watchdog daemon monitors system health every 60s with 3-strike escalation policy.
Shown: first-generation 5-keypoint model. Current pipeline uses the 16-keypoint schema described below.
YOLOv8-pose model trained on 16 anatomical keypoints: 9 Tier 1 body-axis landmarks (snout, eye, gill slits, pectoral base, dorsal base, caudal) and 7 Tier 2 fin tips, for white shark body condition assessment from underwater video. Extracts angle-tolerant morphometric ratios including total length, fork length, head-to-total ratio, and body condition index across 4 California research sites.
Semi-supervised 8-stage pipeline: video cataloging → frame extraction → seed annotation → model training → active learning selection → human correction → pseudo-label amplification → refinement. Processes 4,850 videos (661 GB) spanning 2012–2026 across Año Nuevo, Aptos, Farallon Islands, and Point Reyes. Uncertainty + k-center diversity scoring selects the most informative frames for annotation, with pre-annotation from the current best model to accelerate human review.
Presented research at multiple conferences including the CSUMB Summer Symposium on ML-based dorsal fin morphometrics for white shark identification and health assessment.
Automated harbor porpoise individual identification from dorsal fin photos. SAM 2 segments the body, protrusion geometry extracts the fin, MegaDescriptor-L-384, chosen over DINOv2-base and a two-backbone ensemble on this catalogue, embeds the crop through an ArcFace head, and cosine similarity ranks the 198-individual catalog.
Leave-one-out evaluation on individuals with ≥5 sightings yields 93.0% rank-1 and 98.3% rank-5 accuracy. Temporal split (train ≤2022, test 2023+) gives 30.5% rank-1, which reflects the real challenge of fin shape change over years. Gallery size experiments confirm more photos per individual directly improve re-ID accuracy.
Gradio web interface for field researchers to upload dorsal fin photos and receive ranked identity matches. Includes catalog browsing with individual sighting histories, Dropbox integration for photo ingestion, and batch processing. Deployed on an 8-node Raspberry Pi Kubernetes cluster processing 360 images/hour.
Full-stack Flask annotation platform for shark scar and morphometric data. SAM2 AI segmentation for body and scar detection, 15-point keypoint skeleton with auto-generation (click snout + caudal, remaining 13 points placed automatically), and 12-zone body diagram with 11 scar types. Multi-annotator consensus scoring with weighted voting by experience, proficiency, and confidence tiers.
Interactive 12-zone SVG body diagram for precise anatomical scar placement. Each zone is clickable and color-coded. Scar annotations include type classification, confidence rating (1–5 stars), side (left/right/dorsal/ventral), and free-text notes. Auto-save triggers every 5 minutes with unsaved-changes browser guard.
Export pipeline supporting COCO JSON, Google Forms CSV, Keypoint CSV (wide format), and Google Sheets (3-tab: Scars, Keypoints, Admin Flags) with auto-sync on every annotation save. Google Drive integration auto-resolves video IDs with 3-tier resolution (local → cached → Drive). Deployed at annotate.shark-id.org on AWS EC2 with Docker Compose and nginx reverse proxy.
Sequential Monte Carlo particle filter with a forward-filter backward-sample smoother. Dead reckoning from accelerometer and magnetometer data drifts without bound; GPS and acoustic fixes enter as observations and the backward pass redistributes the correction across the whole track rather than snapping it at each fix. Bathymetry enters as a constraint, so a reconstructed track cannot cross land.
A reconstructed track is only useful if it leaves the project. Readers and writers for CF-1.8 NetCDF, Movebank and GeoJSON, so a track can go into a national archive, a movement-ecology repository or a GIS without a bespoke converter. DeepLabCut and SLEAP keypoint exports come in the other direction as optical ground truth.
Splits are taken by animal, never by record. A random split puts the same individual on both sides, and a behavior classifier can then score well by recognising the animal rather than the behavior. Subject-level leave-one-out, with per-class precision, recall and F1 reported next to balanced accuracy and macro-F1 so a majority class cannot hide inside a single headline number.
LlamaIndex RAG pipeline with LanceDB vector store for citation-aware Q&A across shark research papers. Every answer includes [Author et al., Year, p.X] citations traced to source documents. Supports PDF, DOCX, HTML, EPUB, and Markdown ingestion with Docling section-aware parsing and PyMuPDF fallback.
Automated literature discovery via OpenAlex API with Unpaywall PDF downloading. Pre-configured for 20 shark research domains, from morphometrics, telemetry, ecology, machine learning, genetics, behavior, conservation, through conservation, targeting white sharks, leopard sharks, and orcas. Gradio UI with 7 tabs for chat, upload, library management, summarization, and settings.
Performed CFD analysis using OpenFOAM to model hydrodynamic drag around acoustic tags on shark dorsal fins. Iterative simulations refined tag geometry to minimize swimming resistance.
Ten SQL Server migrations with rollback scripts, restructuring annotations into an observation, localization and geometry hierarchy. Several contributors can now attach geometry to the same animal while license, observer and review state stay attached to each contribution rather than to the animal.
The REST API could describe a bounding box and nothing else. Extending it to points and COCO compressed RLE segmentation meant the Java side had to speak the same codec Python tooling writes, so pycocotools' RLE implementation was ported and covered by 17 unit tests. Masks produced by SAM or any COCO tool round trip byte for byte, with no new dependency added to the server.
Converts existing bounding boxes into segmentation masks across the whole database, scaling from a SAM 2 pilot to a full corpus SAM 3 run. Sharded by image over a WAL mode SQLite manifest used as a lock free work queue: workers claim shards with BEGIN IMMEDIATE leases so a dead worker's shards are reclaimed automatically, commits are temp and rename so a crash never marks a half written shard complete, and one long lived worker per GPU keeps the model loaded once.
Nothing is accepted on faith. Each mask takes the best of three multimask candidates by predicted IoU, carries a stability score computed from mask logits, and lands in one of three tiers on dual thresholds. Rejects are routed to a sidecar with an explicit deferred count, so the accept rate cannot be inflated by quietly dropping the hard cases.
The annotation editor shipped with a stub behind its SegmentationProvider seam that returned a disc where a mask should be. Filled with a loopback ONNX service running SlimSAM-77-uniform through ONNX Runtime and CoreML, so Smart Select returns the outline of the actual animal.
Answers what you actually get if you train on a proposed slice of an annotated survey corpus, from bounded SQL aggregates rather than a full export. Reports an exclusion waterfall attributing every dropped annotation to one named stage, per concept trainability thresholds and the untrainable long tail, train and test leakage counts under four candidate split keys, and unlabelled positive contamination.
End-to-end pipeline for TECAN plate reader growth curve analysis on bacteria remediation strains. Automates the full workflow from raw data to publication-ready figures.
Multi-DOF robotic arm exploring kinematics, servo control, and mechanical design.
Custom monitoring devices measuring pH, temperature, DO, and salinity in real-time.
Research platform and learning opportunity for avionics. Custom-built FPV drone serving as a hands-on testbed for flight controller programming, PID tuning, ESC protocols, and radio telemetry systems.
The story behind the science: eight chapters, one thread
"The frontier in marine science is not just about reaching new places, but about seeing what is already there in radically new ways."
Chapter I
Before College
The first time I saw ROV footage of a hydrothermal vent, the shimmering fluid and tube worms crowding a chimney in complete darkness, something shifted. It was during a thesis defense I'd stumbled into, a graduate student describing species that had never been cataloged, ecosystems sustained by chemistry instead of sunlight, DNA analysis revealing evolutionary histories that rewrote textbooks.
That experience showed me that entire ecosystems on our own planet remain unknown, and that the tools we build to observe them determine what we can discover. It is the reason I now build those tools.
