LLM Application Development
AI/ML Advanced 1 year experience
Summary
I build LLM applications end-to-end from prompt design and context assembly to output validation and cost management. My focus is on making LLM outputs reliable and safe in production by grounding answers in retrieved context, validating generated code before execution, and structuring prompts to handle edge cases rather than assuming the happy path.
How I Apply This Skill
- Designed a dual-API routing system (Anthropic Sync vs. Batch API) that automatically selects the processing mode based on job size, delivering 50% cost savings on bulk workloads across 1,028 files
- Debugged a silent data corruption bug in the Batch API integration where per-batch index resets caused file 100’s metadata to overwrite file 0 across batch boundaries; fixed with global indices and caught through progressive scale testing (2 → 6 → 52 → 122 → 782 files)
- Engineered conditional prompt templates for three distinct input states (existing YAML, legacy inline metadata, no metadata), allowing a single pipeline to correctly transform heterogeneous vault content without manual triage
- Built a 5-stage context assembly pipeline for Text-to-SQL that embeds the user’s query, maps it to relevant database concepts, retrieves and ranks schema context from Qdrant, and passes a deduplicated context window to Claude which prevents hallucinated column names
- Implemented a 7-check SQL security validator that intercepts all LLM-generated SQL before execution, blocking DML, UNION injection, system table access, and other attack vectors regardless of what the model produces
- Built a three-model AI architecture for the Job Search Agent: Haiku 4.5 for high-volume enrichment (0.72 each), and Claude Code for self-healing maintenance
- Designed cross-model orchestration for the Copilot Agent Pipeline, assigning 4 models (GPT-4.1, GPT-5.4, Claude Opus 4.6, Claude Sonnet 4.6) to 7 specialized agents calibrated by cognitive load
- Engineered a pluggable LLM backend for the RAG Document Assistant that swaps between Anthropic and OpenAI-compatible providers via a single environment variable with zero code changes
Key Strengths
- Prompt engineering: structured output, conditional templates, format-specific instructions for heterogeneous inputs
- Context window management: relevance ranking, deduplication, and schema enrichment before LLM calls
- Output safety: post-generation validation layers for code and SQL; treating LLM output as untrusted until verified
- Cost optimization: Anthropic Batch API, model selection (Haiku vs. Sonnet by task), and caching to minimize redundant inference
- RAG grounding: source attribution, similarity thresholds, and zero-hallucination design patterns