AI engineer. St. Augustine / Jacksonville, FL.
I ship production systems as one engineer. What makes that work is the harness: a spec-driven agentic SDLC with a skills library, adversarial review gates that can veto a change, and architecture rules enforced in lint instead of written in a wiki. Agents write most of the code. The harness is what makes it safe to keep.
Everything below came out of that loop. Most of the client work is private; these are the public ones.
Written up:
- A solo engineer's production SDLC with AI agents
- Lint as architecture enforcement
- An independent AI code review gate
- Verify every write to the system of record
Practice: evanparra.ai
Employed with a regional commercial electrical contractor, NE Florida. Software engineer building custom apps inside the Azure tenant: forecasting, change orders, approvals, reporting. Their tenant, their repo, their code, and the team ships on the same harness I do.
Products I run:
- TextTimeline: legal document intelligence. Messy text exports become chronological timelines with a citation on every entry. FAISS + BM25 hybrid retrieval, Cloud Run, Firestore, Gemini. (Source private, paid product.)
- GammaRips: overnight options-flow scanner. 14 Cloud Run services, ~20 schedulers, multi-agent ADK publishing layer with deterministic compliance gating.
Trading and data platform
- gammarips-engine: signal platform over ~10GB/day of market data. LLM-augmented ETL, MCP tool server, GitHub Actions to Cloud Build to Cloud Run. Python, BigQuery, Vertex AI, Pub/Sub.
- gammarips-webapp: customer-facing surface. Daily picks, subscription billing, compliance disclosures.
- gammarips-mcp: MCP server so agents can query financial data. FastMCP on Cloud Run, SSE transport.
Generative AI and evaluation
- genai-eval-framework: hallucination detection via cross-encoder NLI plus semantic similarity, content safety scoring, and A/B model comparison with paired t-tests. HTML and JSON reports for CI. Transformers, Sentence-Transformers, Detoxify, Pydantic.
- lora-finetune-lab: QLoRA fine-tuning with 4-bit NF4 quantization, PEFT adapters, TRL SFTTrainer, and W&B tracking. Transformers, PEFT, TRL, Accelerate.
- diffusion-style-transfer: SDXL base and refiner with IP-Adapter style conditioning, CLIP-based consistency scoring, NSFW filtering. Diffusers, OpenCLIP, PyTorch.
- whisper-multimodal-pipeline: audio to transcription to Gemini analysis to Pydantic-validated JSON. Whisper and Google STT backends, async with retries.
Agents and RAG
- healthcare-graph-rag-agent: clinical Q&A over a medical knowledge graph, citation-backed. ADK, Gemini, Spanner Graph, Cloud Run.
- galatiq-invoice-agent: multi-agent invoice lifecycle (ingest, validate, approve, pay) with self-correction on extraction. LangGraph, FastAPI, Cloud Run.
- serverless-pii-vault: event-driven file storage with user isolation and irreversible PII redaction. Cloud DLP, Vertex AI, Cloud Run.
- SciPaper-Chat: multi-document paper Q&A with citation tracking. Vertex AI Vector Search, Gemini, Firestore.
- yolov9-object-detection-guide: end-to-end guide to fine-tuning YOLOv9 on custom datasets. Written during my M.S. coursework. PyTorch.
GenAI: Diffusers, PEFT/LoRA, Whisper, Stable Diffusion, CLIP
ML/AI: Vertex AI, Gemini, PyTorch, TensorFlow, Scikit-Learn
Evaluation: Sentence-Transformers, Detoxify, W&B, custom frameworks
Cloud: GCP (BigQuery, Cloud Run, Pub/Sub, Vertex AI), Azure on client work
MLOps: GitHub Actions, Cloud Build, Docker, model registry
Data: Python, SQL, Pandas, dbt, Airflow
Backend: FastAPI, Python, Node.js
Frontend: Next.js, React, TypeScript
- M.S. Artificial Intelligence, Florida Atlantic University
- B.A. Economics, Florida International University
- Google Professional Machine Learning Engineer
- Google Advanced Data Analytics