Computer Science & Engineering Student @ RSET (Graduating 2027)
Focusing on Software Development Engineering (SDE), AI Infrastructure, Local AI Orchestration & Model Reliability
- β‘ AI Infrastructure & Compute: Experienced in orchestrating remote CUDA instances on Vast.ai (cutting GPU compute costs by 65%) and fine-tuning 32B parameter models on AMD Instinct MI300X hardware via ROCm 6.2.
- π§© Systems & Formal Verification: Building neuro-symbolic AI agents integrating PyTorch with symbolic solvers (Z3 Theorem Prover) to mathematically validate system constraints and catch edge-case failure modes.
- βοΈ Software Engineering: Architecting asynchronous REST APIs in FastAPI, containerized workflows in Docker, and custom multi-agent RL environments complying with Meta PyTorch OpenEnv standards (1,000+ step rollouts/sec).
- π Hackathons: 3rd Place @ HackS'US Edition V | National Round 1 Qualifier @ Meta PyTorch OpenEnv Hackathon.
- Languages & CS Core:
PythonCJavaData Structures & AlgorithmsSystems ProgrammingOOPDBMS - AI Infrastructure & GPU Acceleration:
PyTorchCUDAVast.aiAMD ROCm 6.2DockerLinux Administration - AI & Frameworks:
Multi-Agent RLLLM Fine-Tuning (QLoRA)Gemma Vision APIGemini Flash APIOpenCVFastAPI - Developer Tools & Solvers:
Z3 Theorem ProverGit/GitHubREST APIsSupabasePyMuPDF
βοΈ JurisSim β Neuro-Symbolic Legal Auditor & Verification Agent
Fine-tuning 32B LLM on AMD MI300X with Z3 Theorem Prover integration.
- Solved PyTorch ROCm SDPA attention overflow bugs using eager attention execution and gradient accumulation to reach stable validation loss (1.676) without VRAM OOM crashes.
- Implemented an agentic self-correction loop where execution errors are intercepted, fed back to the LLM, and validated via mathematical proofs (SAT/UNSAT).
π¦ Autonomous Multi-Vehicle Simulation β Braess's Paradox Engine
PyTorch MARL traffic simulation running on remote CUDA clusters.
- Provisioned on-demand remote CUDA hosts on Vast.ai, reducing GPU compute costs by 65% during 10,000+ episode simulation runs.
- Quantified an 18.4% network latency spike under User Equilibrium and validated a dynamic tolling mechanism that restored system throughput.
π€ Nerk β Custom RL Environment (Meta PyTorch OpenEnv)
Modular multi-agent RL environment adhering to Meta OpenEnv standards.
- Optimized state vector representations and step reward calculations, enabling 1,000+ parallel step rollouts per second during policy evaluation.
- Qualified for the national grand finale of the Meta PyTorch OpenEnv Hackathon.
π± DareDay β AI-Verified Social Media Application
Full-stack mobile app with autonomous vision verification.
- Integrated Google's Gemma Vision API to achieve sub-2-second proof verification latency with 92% evaluation accuracy.
- Awarded 3rd Place at HackS'US Edition V out of 30+ competing teams.
- LinkedIn: linkedin.com/in/mark-joseph-0407682a7
- Email: markjoseph2003@gmail.com