Mike Hall

Rails 8 platform for multi-source job ingestion, semantic matching, and application automation, built as a working laboratory for local-first LLM orchestration and treating scraped third-party text as hostile input.

• Data Ingestion & Normalization: Built a multi-source Rails ingestion pipeline and PostgreSQL storage schema, unifying fragmented job market datasets for structured downstream analysis.
• Semantic Search & Matching: Implemented vector search embeddings via pgvector, enabling high-precision semantic matching between complex candidate profiles and job requirements.
• Local-First AI Orchestration: Designed a provider-agnostic LLM orchestration layer behind a YAML-driven model registry, local-first by default (llama.cpp/Ollama primary, hosted Claude as fallback), so routine inference never leaves the machine while a stronger hosted model can be selected per call site.
• Prompt Injection Defense: Built a four-stage guardrails pipeline (normalize, weighted heuristic scan, risk classify, output validate) screening every untrusted scraped document before it reaches a model, scoring known injection patterns plus an instruction-density heuristic to secure ingestion against hostile inputs.
• Browser Extension & Lifecycle Capture: Built an MV3 Chrome extension with 16 provider adapters that extracts postings from live ATS pages and captures the full application lifecycle (questions asked and answers given) back into the platform.
• Static Signal Integration: Injected static analysis rules and architectural constraints into model context windows, preventing drift between generated outputs and system security contracts.

Skills: Ruby on Rails, PostgreSQL, pgvector, Vector Search & Embeddings, LLM Orchestration, Prompt Injection Defense, Local Inference, Chrome Extensions (MV3), API Design, Data Pipelines, System Design, OpenTelemetry, Distributed Systems