Large Language Models: architecture, pre-training, and alignment safety.
I am a deep learning engineer focused on optimizing model capability, safety frameworks, and scaling
efficiency. I research transformer pre-training architectures and develop safety-critical alignment
frameworks to prevent prompt injections and tool exploitation.
Core Areas
Pre-training & Adaptation
Pre-training medical and domain-specific transformers using SwiGLU, Grouped Query Attention, and
Rotary Embeddings. Architecting memory-efficient adaptations using LoRA/QLoRA on cloud clusters.
Alignment & Model Safety
Implementing alignment feedback loops (SFT, DPO, GRPO) for safety and instruction compliance.
Building multimodal input/output guardrails for VLM applications — chain-of-thought safety
classification served at sub-200ms via vLLM, with false-negative-rate diagnostics per policy category.
Experience & Education
Chronology of professional internships and academic coursework.
May 2026 — Present
A10 Networks
San Jose, CA
Deep Learning Engineer Intern — AI Security Team
- Developed a multimodal input/output guardrail system for VLM-powered applications, fine-tuning a compact language model with SFT and chain-of-thought rationales — serving sub-200ms inference on Hopper GPUs via vLLM and outperforming leading public guardrail benchmarks by a wide margin.
- Built a custom multi-stage data annotation and verification pipeline to overcome the scarcity and noise of publicly available multimodal safety datasets, producing a curated training corpus with inter-annotator agreement metrics and automated quality filtering.
- Designed a systematic evaluation framework with false-negative-rate diagnostics across safety policy categories, collaborating within a 3-person AI Security team with weekly syncs and Weights & Biases experiment tracking.
Vision Language Models · SFT · vLLM · AI Safety · Guardrails · W&B
Oct 2025 — Present
Routes Technologies
Remote, TX
AI Engineering Intern
- Architected an LLM-powered NL-to-SQL RAG pipeline with few-shot prompting, 470+ synonym mappings, parameterized SQL sanitization, session memory, and W&B observability — deployed across 5 Azure ML managed endpoints.
- Built SVD collaborative filtering (scikit-surprise) with NDCG@5 / Precision@K evaluation, paired with an LLM-driven taste profiler via nightly ETL.
- Engineered multi-source recipe ingestion: Scrapy, Instagram Graph API, TikTok oEmbed, GPT-4o-mini classification, and faster-whisper audio fallback.
Python · Azure ML · OpenAI API · scikit-surprise · W&B · Scrapy
May 2025 — Aug 2025
Dreamable Inc.
San Francisco, CA
AI Engineering Intern
- Fine-tuned Qwen-2.5-7B using Hugging Face, PyTorch, and LoRA on Lambda Cloud — deployed on GCP Cloud Run.
- Spearheaded dataset curation pipeline using Pandas, NumPy, and HF Datasets for Q&A task optimization.
- Built an AI-powered Outreach Agent using LangChain, Exa.ai, and OpenAI API.
Hugging Face · PyTorch · LoRA · GCP Cloud Run · LangChain · W&B
Education
San Jose State University
2026 — 2027
B.S. Computer Science · GPA 3.94
- NSP research with Professor William
- CodePath Advanced DSA training
- Dean's List Honoree
San Francisco State University
2023 — 2025
Computer Science (Transferred)
- VP of AI Club & Tech Lead at SparkSF
- Hosted SFHacks (400+ attendees)
- Dean's List Honoree
Projects & Models
Pre-trained models, adapters, and open-source repositories.
Repositories
LLM Firewall for Agentic Tool-Calling
Low-latency inline defense intercepting prompt injections. GPT-2 attacker loop reduced bypass from 23.1% → 3.72%. BERT + LoRA adds ~20ms latency.
GitHub ↗
theHelper — AI Research Assistant
Production RAG with FAISS, LangChain chunking, cross-encoder reranking. Local observability into daily JSONL. CI-gated QA.
GitHub ↗
Kanting — Video RAG System
Indexes YouTube transcripts via Whisper. Semantic search across Sentence-Transformers DB returning precise clip timestamps.
GitHub ↗
End-to-End LLM Post-Pretraining
SFT and GRPO policy alignment pipeline on StableLM 1.6B.
GatorGPT
63M transformer for consumer GPUs. Grouped Query Attention and RoPE.
How LLMs Are Made
Annotated code building GPT-2, DeepSeek MoE, and Kimi from scratch.
GitHub ↗
Hugging Face Weights
MedAssistGPT 303M & 401M
Medical-domain transformers on PubMed. SwiGLU, GQA, RoPE.
HuggingFace ↗
Qwen2.5-0.5B SFT+DPO 0.5B
Chat model fine-tuned with SFT and Direct Preference Optimization.
HuggingFace ↗
Llama-3.2-3B OpenHermes 3B
QLoRA on filtered OpenHermes conversational datasets.
HuggingFace ↗
StableLM 1.6B SFT+GRPO 1.6B
Aligned via GRPO on PKU safety preferences.
HuggingFace ↗
Skills & Credentials
Technical expertise, hackathons, and certifications.
LLM Engineering
Transformers · SFT · DPO · GRPO · PPO · LoRA/PEFT · TRL · vLLM · Quantization · Vector DBs · Prompt Eng.
ML & NLP
PyTorch · TensorFlow · Scikit-learn · LangChain · FAISS · Whisper · BART · Sentence Transformers · Pandas · NumPy
Backend & Cloud
FastAPI · Flask · Docker · Azure ML · GCP · PostgreSQL · MongoDB · Scrapy · Nginx · CI/CD
Programming
Python · SQL · Java · JavaScript · C++ · Bash · R · HTML/CSS · Git · Linux
Hackathons
CalHacks 12.0 — Palace of Fine Arts, SFOct 2025
MCP AWS Agentic Challenge — AWS Builder Loft, SFJul 2025
SacHacks — VirtualMar 2025
HackMerced — UC MercedMar 2025
Cal Hacks 11.0 — San FranciscoOct 2024
Certificates
- AI Memory: LLM Memory Systems — LinkedIn
- Fine-Tuning for LLMs: Beginner to Advanced — LinkedIn
- Model Context Protocol (MCP) — LinkedIn
- Introduction to Generative AI — Google Cloud
- Introduction to Web Development — UC Davis
- Programming in Python — University of Michigan
- Special Theory of Relativity — Stanford University
- Calculus through Data & Modelling (×4) — Johns Hopkins
Get in Touch
Send a brief message to open collaboration.
Open to inquiries about custom fine-tuning runs, alignment evaluation, and model safety audits.