Munish Persaud

Hands on, real stakes

Most engineers learn the field they're building for secondhand: interviews, articles, a ride-along. Healthcare doesn't allow that. I learned it at 3 a.m., on the floor at AdventHealth, as a CNA. That's why the AI I build targets healthcare: I've worked inside the workflows, and I know exactly where they break.

The research half of my work is AI safety: stress testing generative models and building the frameworks that make them safer to release. The rest goes into MAIA, the company I'm building.

The thread through all of it: hands on work with real stakes. Workflow efficiency for the small businesses their communities depend on, better leads through custom CRMs, and safer AI for everyone who has to live with it.

Work

Command Post Technologies

• —Project lead and sole AI engineer on an internal CRM with agentic AI. I own the AI work, the backend, and the frontend. Time from lead to pursuit is down 22%, manual data entry down 63%, opportunity review meetings down 19%.
• —I interview the business development team to find what hurt in the old CRM, then build the fix: automated data entry, opportunity ranking on internal metrics, and shared tracking of team pursuits.
• —Backend runs on a company hosted VM with Node.js and PostgreSQL, alongside a Neo4j Graph-RAG knowledge base and a vector DB matching incoming opportunities to company capabilities.
• —Models run on an NVIDIA DGX Spark with vLLM and Docker. Gemma 4 26B-A4B for agent automation and OCR, GPT-OSS 120B for long document summarization and deeper reasoning. Built preprocessing for consistent I/O and guardrails that keep agents inside company policy.

NSF REU, UCF Center for Research in Computer Vision

• —Developed a new method that bypassed Evo2's safety protocols with a 28% higher success rate than the prior state of the art.
• —Designed and built the evaluation framework for testing model vulnerabilities and safety gaps across generative biology use cases.
• —The results made the case for stronger safety protocols and fed into policy recommendations for generative biology.

AdventHealth Fish Memorial

• —Direct patient care on a hospital floor: mobility, vitals, daily living support. Three years of seeing exactly where clinical workflows hold and where they fall apart.
• —I chart patient status and flag changes to the nursing and physician teams: clear notes, clean handoffs.
• —It's the reason I build what I build. I don't have to guess how a hospital actually runs.

Research

Dr. Bedi's SAFERR AI Lab, UCF

• —Led safety testing for Evo2, a 7B parameter generative biology model, and designed the evaluation workflows and testing protocols.
• —Deployed and optimized 3 models on UCF's HPC cluster across up to 4 linked H100s, with containerized environments for reproducible inference.
• —Analyzed results across 100+ test scenarios to surface safety vulnerabilities. I'm first author on the resulting paper, now under review at NeurIPS 2026, proposing safety improvements for these models.

Projects

Agentic AI that automates the administrative grind of private medical practices. Multi-agent orchestration (LangGraph + Light-RAG) cuts prior authorizations from an industry average of ~20 minutes to ~7 per request, validated across 3 pilot tests; a second pipeline turns physician voice notes into finished EHR documentation in 11 minutes per encounter, beating the physician average of 16 minutes. Built on 50+ interviews with doctors and healthcare professionals.

A React Native app for geriatric fall risk assessment, built HIPAA compliant. Pose estimation runs on device at 30 FPS (MediaPipe), scoring 23 Otago exercises plus Chair Rise and Timed Up and Go. A Qwen3.5-2B model, tuned with QLoRA and quantized to 1.2 GB, handles chat fully offline so patient data never leaves the phone. I led a team of five from first sprint to production IT handoff, with the UI in English, Spanish, and Haitian Creole.

A native VisionCamera v4 frame processor for iOS (Swift + Metal) and Android (Kotlin with the GPU delegate), exposing 33 3D pose landmarks and up to 21 hand landmarks per frame. Inference runs around 10 to 20 ms with temporal smoothing built in. I pulled the CV layer out of AI-PEER, generalized it, and published it on npm, now past 230 downloads, so nobody has to write this plugin twice.

Benchmarked 10 video understanding models on Something-Something V2 (174 classes, ~220K clips). Tuned Qwen3.5-4B with QLoRA at 4 bits to 58.19% accuracy while training just 0.12% of the model, landing within 6 points of a frozen V-JEPA ViT-L baseline at 1/100th the trainable parameters. Also reproduced Meta FAIR's V-JEPA eval and traced a gap of 5 points to a hardcoded float16 autocast silently overriding the BF16 config.

VQ-VAE neural compression for whole slide pathology images: 272 MB down to 3.24 MB, a 98.8% reduction, with diagnostic quality preserved. I pitched the idea, led a team of four through the weekend, and validated the clinical need with 7 medical professionals. React frontend, Flask API, real time inference on a cloud L40S GPU.

A preview tool for plastic surgery consults. A Swift iOS app captures facial geometry, a GPU inference server runs DreamOmni2 tuned with LoRA on procedure specific data, and a clinician dashboard shows the predicted outcome. Guardrails built on CLIP restrict inputs and outputs to facial imagery only.

Stack

Honors

1st Place, UCF Innovate Startup Pitch Tournament

Volunteer of the Year