AI Engineer building production systems that ship and stay running.
A library is not a service — What it takes to put an LLM information extractor behind real traffic
Part one found the winning configuration; part two builds the event-driven AWS service around it and puts both extraction strategies under the same 200-document burst. The system trades latency, never cost, for stability. Under load, the agentic queue drains at half the speed, a premium the offline benchmark couldn't see.
MLOps · AWS · Load Testing
When does agency earn its cost? — Benchmarking LLM extraction strategies on legal documents
Same documents, same schema, three model families. The agent rarely beats a single LLM call. A clean counterexample to the reflex of wrapping every task in an agent loop.
LLM Engineering · Agents · Evals
agentic-kie-deploy — Production-grade asynchronous AWS pipeline
Terraform-provisioned, event-driven AWS infrastructure that turns the agentic-kie library into an asynchronous extraction service. A caller asks for an upload slot, pushes a document straight to S3 via a pre-signed URL, and reads the structured result from an address it already holds—everything in between stays decoupled.
Terraform · AWS · API Gateway · Lambda · EventBridge · SQS · DynamoDB · S3
agentic-kie — Schema-driven key information extraction from documents
Typed Python library where a PDF enters as a file path and leaves as a validated Pydantic instance. Handles text-layer detection, OCR routing, image rendering, LLM orchestration, and retry logic. Two extraction strategies—single-pass and agentic ReAct—satisfy the same protocol and swap without touching downstream code. Model-agnostic via LangChain's BaseChatModel.
Python · LangChain · Pydantic · ReAct
📬 Writing about Software Engineering, MLOps and LLM Engineering → gabriel.com.gt/blog