Sahiel Bose
UCLA CS + Math
I study Math + CS at UCLA. I'm an Undergraduate Researcher at Yale in the Gerstein Lab, working on CellForge, a multi-agent AI scientist system for single-cell analysis, and at Carnegie Mellon University, developing longitudinal segmentation for serial brain MRI to quantify cortical layer atrophy in Alzheimer's disease, and a Research Software Engineer at Johns Hopkins, where I'm integrating 3D interactive segmentation into BodyMaps, a CS and Medicine program building a body atlas for early cancer detection.
I'm also a Research Engineer at ThinkNeuro, working on real-time machine learning that interfaces with medical sensors to detect opioid overdose events. Alongside that, I've spent the last couple of years collaborating on academic research regarding logical reasoning in LLMs and graph neural networks with mentors at UC Santa Cruz and UC Riverside. From May to October 2026, I was a Founding Engineer at OpenSwarm, an Entrepreneur First-backed agentic browser, where I built end-to-end agentic workflows and the system that automates social media across five platforms.
Contact
Education
B.S. Mathematics of Computation, UCLA
Experience
Undergraduate Researcher, Yale University
Oct 2026 – Present | New Haven, CT
Working on CellForge, a multi-agent AI scientist system that designs single-cell analysis models, in the Gerstein Lab.
- Working with Daniel Shao in the Gerstein Lab on CellForge, a multi-agent AI scientist system that designs single-cell analysis models, running parallel GPU experiments for the paper's revision.
- Analyzing agent reasoning traces from CellForge runs and prototyping a more flexible agent workflow to replace its hand-coded, staged pipeline, so agents can plan and revise their own analysis steps.
Undergraduate Researcher, Carnegie Mellon University
Sep 2026 – Present | Pittsburgh, PA
Developing longitudinal segmentation for serial brain MRI to quantify cortical layer atrophy in Alzheimer's disease.
- Developing longitudinal segmentation for serial brain MRI in the Xu Lab to quantify cortical layer atrophy in Alzheimer's disease, improving on independent per-timepoint masks without ground-truth labels.
- Reviewing longitudinal segmentation methods and designing evaluations that test measured atrophy against expected disease trends and compare unregistered with registered scans.
Research Software Engineer, Johns Hopkins University
Aug 2026 – Present | Baltimore, MD
Integrating 3D interactive segmentation into BodyMaps, a Johns Hopkins CS and Medicine program building a body atlas for early cancer detection.
- Integrating nnInteractive, a 3D interactive segmentation framework, into BodyMaps, Prof. Zongwei Zhou's medical-AI program (Johns Hopkins CS and Medicine), for in-browser CT structure segmentation.
- Fine-tuning the segmentation model on Johns Hopkins Medicine CT data and benchmarking it on the PanTS dataset to improve per-structure accuracy for the 3D body atlas built for early cancer detection.
BodyMaps on GitHub
Research Engineer, ThinkNeuro LLC
Nov 2024 – Present | Sacramento, CA
Leading ML for a patent-pending wearable that detects opioid overdose in real time, trained on a 721K-sample dataset.
- Leading ML integration for a nonprovisional patent-pending wearable for real-time opioid overdose detection; built GRU, LSTM, and logistic-regression classifiers on a 721K-sample dataset with subject-level splits.
- Engineered the biphasic-overdose feature set and co-authored the research paper and literature review; built the React Native app that streams live vitals off a four-sensor Arduino Nano ESP32 array and alerts caregivers.
Repository
Founding Engineer, OpenSwarm
May 2026 – Oct 2026 | San Francisco, CA
Building agentic workflows and cross-platform social automation on an Entrepreneur First-backed agentic browser with 5k+ active users.
- Developing OpenSwarm, an Entrepreneur First-backed agentic browser with 5k+ active users: built a system that automates social media across 5 platforms, handling scheduled posting, engagement, and analytics.
- Author end-to-end agentic workflows on OpenSwarm spanning creative generation, media tagging, CRM automation, architecture mapping, and scheduled overnight jobs and briefings.
Research Assistant, UC Santa Cruz, AIEA Lab
Oct 2024 – Apr 2026 | Santa Cruz, CA
Researching logical reasoning in LLMs on ProSLM, the lab's logic-augmented LLM platform.
- Completed a 10-month onboarding in LLM logic and reasoning (symbolic inference, AI explainability) across Python and Prolog, then joined active research on ProSLM, the lab's logic-augmented LLM platform.
- Built natural-language-to-Prolog pipelines in SWI-Prolog and a from-scratch backward-chaining inference engine, and designed chain-of-thought scaffolds with self-consistency to refine LLM reasoning.
Repository
Research Assistant, UC Riverside
May 2025 – Jan 2026 | Riverside, CA
GNN-based traffic congestion hotspot detection across large-scale road networks, benchmarking four graph-sampling architectures.
- Assisting a PhD researcher in Prof. Amr Magdy's lab on GNN-based traffic-congestion-hotspot detection across large-scale, heterogeneous road networks (intersection and event nodes), applying spatial analytics to real-time geospatial data.
- Trained and benchmarked graph-sampling methods (GraphSAGE, GraphSAINT, GRAPES, HMSG) and heterogeneous-graph architectures to improve scalability and predictive accuracy for congestion prediction.
Projects
Optivia
Agentic Workflows, LangGraph | Mar – Aug '26
A macOS desktop app that decomposes a project into a dependency-aware task graph and staffs each node with a scoped, model-tiered agent, running them in build order with verified context passed downstream.
- Built end to end: Electron and React over a typed IPC bridge to a Python/LangGraph engine, with the agent fleet running on a live streamable desktop.
- Verification gates block any node that fails before it advances, so a bad result never propagates downstream.
Repository
Vigil
Finalist, Anthropic × Abridge | Jul '26
Named a Finalist, top 6 of 115 teams at the Anthropic × Abridge hackathon, for a continuous ER waiting-room re-triage agent that catches patient deterioration between check-in and the exam room.
- A YOLO11-pose model tracks skeletons only for privacy, with fall and seizure detectors I tuned on real clips to cut false alarms.
- An ESI algorithm scores each patient's acuity at intake, then dynamically re-scores it against their chart as events fire, clamped so urgency only rises, before triggering a nurse call and an auto-written SOAP note and FHIR bundle.
Repository
SixthSense
2nd Place, Qualcomm × Meta | Jun '26
Won 2nd Place at the Qualcomm × Meta hackathon for a wearable that guides blind and low-vision users, running four models on a Galaxy S25 camera feed entirely on-device.
- Four models run on-device via ExecuTorch (YOLOv11n, Depth-Anything-V2, Qwen2.5-0.5B-Instruct, ML Kit OCR) against a live Galaxy S25 camera feed.
- Built the vision-to-haptics link: a BLE client and the mapper that converts per-zone depth into directional motor cues, driving an Arduino ESP32 and vibration motors on the harness.
Repository
HealthFlow
1st Place, Apify × Scalekit | May '26
Won 1st Place at the Apify × Scalekit hackathon for a 9-agent system that takes an emergency call all the way to physician-approved orders.
- Four role-specific CRMs on Next.js (911 dispatch, paramedic voice capture, nurse triage, physician order review) carry a single case end to end.
- Integrated Apify-powered drug-interaction and allergy screening with Scalekit authentication and a SHA-256 immutable audit trail for end-to-end clinical safety and compliance.
Repository
SignSpeak
2nd Place, ACSEF | Dec '24 – Mar '25
Earned 2nd Place at the Alameda County Science & Engineering Fair (ACSEF) for an AI-powered ASL recognition system built for two-way, hands-free conversation, with models reaching 95% validation accuracy.
- Co-led the model work: MobileNetV2 transfer learning for static signs and a bidirectional LSTM for dynamic gestures, reaching 95% validation accuracy.
- Deployed on a Raspberry Pi for real-time inference.
Repository
Skills
Machine Learning & Perception
Python, PyTorch, TensorFlow, scikit-learn, YOLO, OpenCV, LSTM & GRU, Transfer Learning, Graph Neural Networks
Agentic & LLM Systems
Multi-Agent Orchestration, LangGraph, LangChain, RAG, LLM Reasoning, Task Decomposition, Model Routing, Prolog, Voice Agents, Browser Automation