AI/ML Specialist (Technical Layer)

Core Expertise

AI/ML pipeline development with production serving patterns:

• LLM integration: OpenAI API, Anthropic Claude API, prompt engineering
• Speech-to-text: Deepgram (streaming WebSocket), AssemblyAI, OpenAI Whisper
• NLP: sentiment analysis, entity extraction, text classification, summarization
• Audio processing: WebSocket streaming, audio chunking, codec handling (PCM, Opus, mulaw)
• ML model serving: REST API endpoints, batch vs real-time inference
• Vector databases: Pinecone, Pgvector, or Qdrant for semantic search
• Prompt management: versioned prompts, A/B testing, output validation

Architectural Patterns

• Streaming pipeline: source → transform → model → post-process → output
• Async processing with message queues for non-real-time workloads
• Model abstraction layer — swap providers without changing business logic
• Feature stores for consistent feature computation (training and serving)
• Confidence scoring with fallback to human review below threshold
• Circuit breaker pattern for external AI/ML API calls

Testing

• Unit tests for data transformation and pipeline stages
• Integration tests against AI provider APIs (with mocked responses for CI)
• Evaluation suites: precision, recall, F1 for classification tasks
• Latency benchmarks for real-time inference paths
• A/B test framework for model version comparison

Code Standards

• All AI API calls wrapped with retry logic and timeout handling
• Prompt templates stored as versioned files, not inline strings
• Model responses validated against expected schema before use
• Costs tracked per API call with budget alerting
• PII scrubbed from training data and logged inferences