| # Chris Austin - Senior AI Engineer |
|
|
| ## Overview |
| Chris Austin is a Senior AI Engineer with 7+ years of experience building scalable data systems and AI solutions, from ML pipelines to multi-agent GenAI apps. Experienced in Python, Spark, SQL, and deploying across Azure, AWS, and hybrid environments. |
|
|
| ## Current Role |
| Senior AI Engineer at Cognizant. |
|
|
| ## Location |
| Los Angeles / Remote |
|
|
| ## Education |
| B.S. in Mathematics and Computer Science from Temple University. |
|
|
| ## Contact |
| - Email: ceaustin117@gmail.com |
| - LinkedIn: linkedin.com/in/christopher-austin |
| - GitHub: github.com/ceaustin117 |
|
|
| ## Featured Projects |
|
|
| ### Multi-Agent GenAI Tax Automation |
| Designed a multi-agent GenAI architecture on Azure that automated tax document workflows, cutting review time from 14 days to 1 day, and enabling real-time AI approval through a React web app. |
| Technologies: Azure, GenAI, Multi-Agent Systems, React, Python |
|
|
| ### LLM vs Rule-Based Classification |
| Built a classification pipeline that compares LLM vs rule-based outputs to cut manual review time by 40% and reached over 95% accuracy. |
| Technologies: LLM, Python, ML Pipelines, Azure |
|
|
| ### Microsoft Fabric Analytics Solution |
| Delivered a Microsoft Fabric analytics solution with semantic search and AI insights, improving decision-making speed and accuracy across operational sources. |
| Technologies: Microsoft Fabric, Azure, Semantic Search, AI |
|
|
| ### Databricks MLOps Pipeline |
| Rebuilt legacy ML pipelines within a modern Databricks MLOps lifecycle framework (MLflow, CI/CD), adding LLM-based monitoring and improving deployment speed. |
| Technologies: Databricks, MLflow, CI/CD, Python |
|
|
| ### Cloud ML Demand Forecasting |
| Built a cloud ML pipeline for demand forecasting, cutting modeling time from hours to minutes and delivering forecasts via a React dashboard. |
| Technologies: AWS, ML Pipelines, React, CI/CD |
|
|
| ### Predictive Vehicle Mapping Pipeline |
| Engineered a predictive mapping pipeline using Azure Data Factory and Python, integrating ML models that improved processing speed by 500% across millions of vehicles. |
| Technologies: Azure Data Factory, Python, ML, Spark |
|
|
| ## Key Achievements |
| - Reduced tax document review time from 14 days to 1 day |
| - Achieved 500% processing speed improvement on vehicle mapping pipeline |
| - Reached 95%+ classification accuracy on LLM pipeline |
| - 7+ years of professional experience in data and AI |
|
|
| ## Skills |
| - GenAI & LLMs: LangChain, Multi-Agent Systems, RAG, Prompt Engineering |
| - ML Engineering: ML Pipelines, MLflow, Forecasting, Classification |
| - Cloud: Azure (Data Factory, ML, Functions, Fabric), AWS, Databricks, Snowflake |
| - Programming: Python, SQL, Spark/PySpark, TypeScript, JavaScript |
|
|
| ## Certifications |
| - Microsoft Azure AI Engineer |
| - Microsoft Azure Fabric Data Engineer |
| - AWS Certified Cloud Practitioner |
| - Databricks Generative AI Engineer |
|
|
| ## This Website's RAG Chatbot Architecture |
|
|
| This chatbot you're talking to is a RAG (Retrieval-Augmented Generation) system built by Chris Austin to demonstrate AI engineering skills. |
|
|
| ### Architecture Overview |
| The system consists of two main components: |
| 1. **Frontend**: React TypeScript app hosted on GitHub Pages |
| 2. **Backend**: Python FastAPI server hosted on HuggingFace Spaces |
|
|
| ### How It Works |
| 1. User types a question in the chat interface |
| 2. Frontend sends the question to the backend API |
| 3. Backend embeds the question using sentence-transformers (all-MiniLM-L6-v2) |
| 4. Cosine similarity search finds the most relevant chunks from the knowledge base |
| 5. Top 3 relevant chunks are sent as context to the LLM |
| 6. Groq API (llama-3.1-8b-instant) generates a response based on the context |
| 7. Response is returned to the frontend and displayed |
|
|
| ### Tech Stack |
| - **Frontend**: React 19, TypeScript, CSS, GitHub Pages |
| - **Backend**: Python, FastAPI, uvicorn |
| - **Embeddings**: sentence-transformers (all-MiniLM-L6-v2 model) |
| - **Vector Store**: In-memory with pre-computed embeddings stored in JSON |
| - **LLM**: Groq API with Llama 3.1 8B Instant model |
| - **Hosting**: HuggingFace Spaces (backend), GitHub Pages (frontend) |
|
|
| ### Why This Architecture |
| - **Cost**: $0 - uses free tiers of Groq and HuggingFace |
| - **Simplicity**: No complex vector database needed for small knowledge base |
| - **Speed**: Groq provides fast inference, sentence-transformers are lightweight |
| - **Customizable**: All code is custom Python, easy to modify and extend |
|
|
| ### Key Design Decisions |
| - Pre-computed embeddings: Knowledge base is small (~6 chunks), so embeddings are computed once and stored in JSON |
| - In-memory search: Simple numpy cosine similarity, no need for Pinecone/Chroma |
| - Groq over OpenAI: Faster and has generous free tier |
| - Separate frontend/backend: Clean separation, backend can be reused for other interfaces |
|
|
