Building production RAG pipelines, LLM systems, and self-hosted AI infrastructure — from model to deployment.
Quick nav: About · Batcave · Projects · Highlights · Blog · Resume
Now: Fine-tuning local LLMs on Batcave · scaling the RAG podcast pipeline · planning the next Home Server Chronicles entry.
I'm an ML/AI Engineer & Data Scientist with 5 years building production ML systems, RAG pipelines, and self-hosted AI infrastructure. I design end-to-end solutions — from model training and LLM fine-tuning to cloud deployment and real-time monitoring.
When I'm not shipping ML at work, I'm operating Batcave — my solo-built, 56-container private AI server running local LLMs, RAG, and ML inference at 99.9% uptime (live status).
role: Data Scientist — GenAI & ML @ Enigma Technologies
education: M.S. Data Science, UMBC (GPA 3.91 / 4.0)
focus: LLMs · RAG · MLOps · Self-hosted infrastructure
homelab: 56 containers · 36+ services · 99.9% uptime
based_in: Maryland, USA- Full-time ML/AI or MLOps engineering roles.
- Contract and advisory work for production RAG/LLM systems.
- Collaborations around self-hosted AI, homelab observability, and applied GenAI.
A solo-managed homelab on a Beelink mini-PC. Self-hosted from scratch. Live at jay739.dev.
| Metric | Value | Metric | Value |
|---|---|---|---|
| Containers | 56 |
Self-hosted Services | 36+ |
| Uptime | 99.9% |
Local LLMs | 5+ models |
| Hybrid-cloud Latency | −60% |
Multi-DB Layer | 177 GB |
Stack: Docker · Ollama · LangChain · Authentik SSO · Tailscale · Netdata · Nginx · PostgreSQL · MariaDB · Redis · Meilisearch
Proof links: Live Batcave status · Batcave architecture series · OCI migration + latency notes
→ Read the full 5-part Batcave blog series
Core: Python · PyTorch · LangChain · Ollama · Docker · AWS · PostgreSQL · Redis · Next.js · TypeScript
Full badge stack
ML / AI
MLOps · Infra · Cloud
Data
Frontend & Edge
|
Solo-built private AI infra on a Beelink mini-PC. 56 containers, 36+ services, local LLMs via Ollama, RAG pipelines, SSO, full observability.
|
PDF → LLM → TTS pipeline that converts books into character-voiced podcasts. 98% OCR accuracy, 4× throughput via LoRA fine-tuning.
|
|
PySpark ML pipeline on IBM's 32M-row HI-Medium AML dataset. Temporal feature engineering, severe class imbalance (0.23%) — 0.998 F1.
|
Benchmarked 8 CNN architectures on the FLAME dataset using RGB+IR dual-stream fusion. Best model: ResNet18 at 0.94 micro-F1.
|
|
Compared custom CNN vs ResNet18 / VGG16 / EfficientNet on 220k+ pathology patches. Transfer learning best at 0.94 AUC.
|
FinBERT sentiment + Monte Carlo VaR across S&P 500 sectors and historical crisis periods (2008, COVID-19).
|
→ See more projects on jay739.dev
More: coding habits & topics
• Benchmarking Local LLMs Across an RTX 3060 Ti and an M4 Mac Mini (With a Kernel Panic Along the Way) Jul 16, 2026
• The Freeze Wasn't Memory: Tracing a Homelab Server's Root Cause Through Three Wrong Turns Jul 16, 2026
• I Made My AI Ambush Me With Pop Quizzes May 29, 2026
• The Quiet Math of Lifetime Software Deals May 12, 2026
• What Running a Homelab Actually Teaches You That Work Doesn't Apr 24, 2026
→ More at jay739.dev/blog · auto-refreshed every 6 hours from jay739.dev/rss.xml
2025 → now Data Scientist — GenAI & ML @ Enigma Technologies
2025 AI/ML Programming Intern @ R/SEEK · UMBC
2023 → 2025 M.S. Data Science (GPA 3.91) @ UMBC
2019 → 2023 Machine Learning Engineer @ Cognizant
2019 Software Engineering Intern @ Infosys