Computer Science + Biomedical Sciences
MUNISH PERSAUD
Building safer AI systems at the intersection of machine learning and healthcare
I believe that as AI systems become more powerful, ensuring their safety and alignment with human values becomes paramount. My work focuses on developing rigorous testing methodologies for bio-foundational AI while creating healthcare applications that genuinely empower patients and practitioners.
FOCUS AREAS
Specialized expertise across AI safety, computer vision, and biomedical applications
AI Safety
Developing adversarial testing frameworks and red-teaming methodologies for generative bio-foundational models. Specializing in vulnerability assessment, jailbreak detection, and safety protocol evaluation.
Computer Vision
Implementing pose estimation pipelines, gait analysis algorithms, and real-time inference systems optimized for edge deployment on resource-constrained hardware.
Healthcare AI
Architecting HIPAA-compliant ML systems with on-device inference, federated learning considerations, and clinical workflow integration for diagnostic and therapeutic applications.
Full Stack Development
Building scalable microservice architectures with React/React Native frontends, Node.js backends, and cloud infrastructure leveraging containerization and CI/CD pipelines.
WORK EXPERIENCE
Industry experience building production AI systems
AI Solutions Software Engineer Intern
Command Post Technologies
- Built agentic AI workflows in LangGraph with multi-agent orchestration, tool use, and bounded reasoning loops for analyst-facing tasks, designing the control flow to run end-to-end inside air-gapped, on-prem environments with no external network dependency
- Deployed LLM-based inference pipelines using vLLM, Ollama, and llama.cpp inside hardened Docker containers on isolated GPU workstations, packaging models, weights, and dependencies into reproducible offline images suitable for DoD and federal deployment targets
- Fine-tuned 4-8B parameter LLMs with QLoRA and applied post-training quantization (Q4_K_M, AWQ) for constrained edge hardware, cutting memory footprint up to 75% and enabling on-device inference on NVIDIA Jetson-class and laptop-grade GPUs
- Built a reproducible evaluation harness benchmarking agent task accuracy, tool-call success rate, p50/p95 latency, and tokens/sec across model and hardware configurations, surfacing regressions before deployment and informing model selection for production pipelines
- Authored deployment runbooks, container configurations, and structured experiment logs for the AI Solutions team; debugged issues spanning model behavior, REST API layer, and CUDA/driver-level GPU configuration to shorten review cycles with senior engineers
FEATURED PROJECTS
Healthcare-focused AI systems designed for real-world impact
AI-PEER
Project Manager / Lead Engineer
HIPAA-compliant React Native app pairing on-device computer vision with a small language model for geriatric gait assessment and fall-risk reduction. Led a 5-person team through a ~50-patient clinic pilot. Ships 23 Otago-protocol exercises and clinical assessments (Chair Rise, Timed Up and Go) using MediaPipe Pose + Hand Landmarker plugins running at 30 FPS on-device, with tuned geometric heuristics for real-time form feedback. Qwen3.5-2B fine-tuned via QLoRA on a 10,800-row geriatric dataset, quantized to 1.2 GB (Q4_K_M), and served through llama.rn for fully offline chat. Google Cloud backend (Cloud Run, Firestore, Firebase Auth) with scheduled REDCap sync. Trilingual UI (English, Spanish, Haitian Creole) with native TTS.
Watch Demo MAIA
Co-Founder / Full Stack Developer / AI Engineer
AI-powered platform automating administrative workflows for private medical practices. Validated product-market fit through 50+ physician interviews across multiple specialties. Currently in active development.
LISA
Founder / Full Stack Developer / AI Engineer
Neural compression system for whole slide images (WSIs) addressing the storage bottleneck in digital pathology where single slides exceed 1GB. Implemented a VQ-VAE architecture with a learned discrete codebook for latent space quantization, paired with a separate decoder network to preserve diagnostic-quality reconstruction. Achieved 84x compression ratio (272.52MB to 3.24MB) with minimal perceptual loss. Built in 36 hours at Startup Weekend Orlando. React frontend communicates with a Flask API backend running inference on a vast.ai L40S GPU instance.
Watch Demo AURA
Founder / Full Stack Developer / AI Engineer
Multi-platform surgical outcome prediction system. iOS client built in Swift captures facial geometry and transmits to GPU inference server. Core generation model is DreamOmni2 fine-tuned with LoRA adapters on procedure-specific datasets to predict post-operative facial morphology. React webapp integrates Web Speech API for voice-controlled surgeon interaction. Implemented content filtering guardrails using CLIP embeddings to restrict input/output to facial imagery only.
FOMO NOMO
Frontend Mobile Developer
Cross-platform mobile application aggregating student organization events into a unified feed. Translated high-fidelity Figma prototypes into production Flutter code with custom widget composition. Implemented infinite scroll pagination with lazy loading for optimized memory footprint. Achieved WCAG compliance through Lighthouse audits, semantic markup, and VoiceOver/TalkBack screen reader testing. Maintained platform parity across iOS and Android rendering pipelines.
RESEARCH EXPERIENCE
Academic positions driving innovation in AI safety and computer vision
Research Assistant
Dr. Bedi's SAFERR AI Lab, UCF
- Lead red-teaming initiative for Evo2, a 7B parameter generative bio-foundational model, designing systematic adversarial testing workflows
- Deploy and benchmark transformer architectures on UCF's HPC cluster using SLURM job scheduling and multi-GPU parallelization
- Develop automated evaluation pipelines to quantify safety vulnerabilities, measuring attack success rates and model refusal behaviors
Undergraduate Research Assistant
Center for Research in Computer Vision REU, UCF
- Executed 100+ adversarial test scenarios against Evo2 bio-foundation model, establishing empirical evidence for inadequate safety guardrails in generative biology
- Developed novel jailbreaking methodology achieving 28% higher bypass rate than existing state-of-the-art prompt injection techniques
- Architected comprehensive evaluation framework with custom metrics for assessing dual-use potential and biosecurity risk in foundation models
- Continuing post-REU research to formalize safety improvement proposals for bio-foundation AI alignment
TECHNICAL SKILLS
Languages
Frameworks
ML / AI
Platforms
Bioinformatics
LET'S CONNECT
Open to research collaborations, opportunities in AI safety, and healthcare AI projects