GPU inference benchmarking framework for studying when speculative decoding helps or hurts LLM inference.

Focus: Speculative decoding | GPU latency benchmarking | Model routing | Reproducible ML systems evaluation

• Benchmarked Qwen 2.5 and Llama 3 draft/verifier pairs on an NVIDIA GH200 480GB.
• Measured speculative-decoding latency across 1,800 benchmark rows and two seeds per model family.
• Characterized slowdown regimes where draft-model overhead can dominate verifier-call savings.
• Built reproducible tooling with raw traces, YAML configs, summaries, tests, and report-style analysis.

LLM document-agent evaluation infrastructure for testing extraction systems under embedded prompt-injection attacks.

Focus: LLM evaluation | Document AI | Agent safety | Provider benchmarking | Reproducible pipelines

• Evaluated OpenAI, Mistral, and Gemini document agents across synthetic receipts, invoices, policies, and bank statements.
• Produced a reproducible 2,700-row benchmark across 300 examples, 3 providers, and 3 prompting strategies.
• Measured malicious-instruction compliance, safe completion, and unsafe tool-execution risk.
• Built a Dockerized FastAPI pipeline with provider abstraction, SQLite caching, structured logging, and resumable runs.

AI-assisted receipt and expense-splitting app focused on making shared payments easier and more accurate.

Focus: Mobile engineering | OCR/Document AI | Expense splitting | Group balances | Product correctness

• Built receipt and statement workflows for scanning, parsing, reviewing, and splitting shared expenses.
• Integrated OCR/LLM-based extraction with validation for items, subtotal, tax, tip, discounts, and totals.
• Designed group balance and settlement flows to help users track who owes whom across shared expenses.
• Focused on correctness and reducing the risk of incorrect payment splits before users send payments.