Software Engineer Intern (AI Development)
ACS (Affiliated Construction Services)
Developed an LLM-powered assistant supporting internal teams and external partners with project-specific inquiries, code and script generation, and contextual guidance.
Python
Azure Government Cloud
Hugging Face
Ollama
RAG pipelines
Embeddings
Prompt engineering
Structured outputs
Agentic system design
Model quantization
- Built end-to-end LLM application features within a cross-functional professional team, integrating with the company's Azure Government Cloud model endpoint
- Designed agentic components using embeddings, RAG pipelines, and NLP techniques
- Led prompt engineering and system design work to improve the assistant's accuracy, scalability, and adaptability
- Studied and applied model internals: tokenization, transformer architectures, fine-tuning, alignment, and optimization techniques including quantization and compression
Deep dive: Reliable structured output
The assistant needed to generate syntactically correct DSL commands. Early iterations validated model output with regex, which broke down on commands with variable-length or customizable inputs that patterns could not fully capture.
The solution was designing structured output formats that constrain the model's generation space at inference time, enforcing schema compliance up front instead of patching malformed output after the fact. The result was a major improvement in parse reliability and a clean separation between the generation and execution layers.