AI Engineer · Los Angeles · UTC−8
Sergio Avila
I build agents and LLM systems — retrieval, evaluation, and inference infrastructure — and ship the production software around them.
Public commit activity, pulled live from GitHub’s contribution calendar.
2,784 contributions · last 12 months
LLM gateway and control plane — virtual keys, atomic budgets, provider failover, per-request cost. And the measurement almost nobody publishes: at the default threshold, a semantic cache answered 98 of 130 unseen questions from a neighbour — 92 of them wrong.
Self-hosted autonomous entity — drives, persistent memory, a dreaming background mind — instrumented on real hardware and debugged at the behavioral level. 1,296 tests.
Agents & tool use
Multi-step agents on foundation models — tool calling, persistent goal state, and the safety rails that keep them supervised.
orchestration · function calling · directive channels · guardrails
Retrieval
Grounded generation with provenance, tuned on real corpora rather than toy benchmarks.
chunking · embeddings · vector stores · cross-encoder reranking
Inference & serving
Local and containerized model serving under real load, on hardware I own and measure.
vLLM · GPU topology · quantization · throughput benchmarking
Evaluation & red-teaming
Harnesses that prove whether a system actually improved — including adversarial coverage that finds the failures polite testing misses.
LLM-as-judge · pairwise evals · red-team suites · failure-rate measurement
Production systems
The software around the model: multi-tenant backends, payments, shipped App Store apps. Products, not demos.
Postgres · RLS · Stripe · iOS · CI/CD
Everyone ships semantic caching. Almost nobody measures how often it silently returns the wrong answer — because measuring that needs exact ground truth. I had some. At the industry-default threshold, one cache hit in five was a wrong answer served with confidence.
I deployed my agent platform to AWS and re-ran its 133-question eval suite as a pre-registered experiment. The scores landed 0.8 points apart. Then the regression gate failed both runs — local and cloud, on completely different categories — and that turned out to be the best result on the table.
I built a 133-question eval suite with an exact answer key and pointed it at my own agent platform. In its first 24 hours it caught four bugs — every one of them mine, none of them the model's. Then it settled which model is actually worth the money.
05Contact
Targeting remote AI Engineer, Applied AI, and MTS roles — building agents, retrieval, evaluation, and inference infrastructure on foundation models, and shipping the production systems around them. Based in Los Angeles.