ABOUT

Trevor Santiago

Applied AI engineer

I'm Trevor Santiago, a data scientist at ALLDATA building AI applications and data workflows. My recent work includes Gemini-powered services, semantic retrieval, and tools that help subject-matter experts reuse prior edits and review AI-generated drafts.

I like taking an unclear problem through research, prototyping, and implementation. At ALLDATA, that has meant working directly with subject-matter experts, building FastAPI services on Google Cloud, and checking model output against the cases those experts care about.

Outside work, I built NormFlow, a text-normalization workbench that combines exact matching, semantic retrieval, and optional LLM suggestions with batch processing and human review.

EXPERIENCE

Experience

A record of roles, responsibilities, and the work behind them.

PROJECT EVIDENCE

NormFlow

Independent project

A text-normalization workbench with exact matching, semantic retrieval, optional LLM fallback, and human review.

• Python
• FastAPI
• TypeScript
• SQLite
• FAISS
• Semantic retrieval
• Human review

I built and shipped NormFlow using Python, FastAPI, TypeScript, SQLite, and FAISS. It combines exact matching and semantic retrieval with optional LLM fallback.

Durable batch, review, and export workflows keep human review and correction within the process of turning raw text into reusable normalized data.

PROJECT EVIDENCE

AI-assisted PCD editing

ALLDATA

An editing workflow that fills from previously applied edits, then uses a separately fine-tuned model to generate new draft edits when matching finds no reusable edit.

• Python
• Semantic retrieval
• Flan-T5
• QLoRA
• BLEU
• Cloud Run Jobs
• Cloud Storage

Probable Cause Data editing turns parsed OEM repair text into curated probable causes. Subject-matter experts used an Excel macro and a library mapping source text to previously applied edits. Its limited matching left unmatched text for manual drafting.

I built a workflow that looks for each input in that library, progressing through exact, case-insensitive, whitespace-normalized, formatting-normalized, and semantic matching. A match fills the corresponding prior edit into the output. Normalization preserves symbols with automotive meaning and exposes conflicting library mappings for expert review.

Inputs without a reusable match go to a separate generation step. This step uses a fine-tuned Flan-T5 Small model to generate new draft edits.

Matching and generation run as separate Cloud Run Jobs connected by Cloud Storage upload events. I also built a simple upload/download UI for non-technical experts. I designed and implemented the system independently apart from the UI's Kubernetes deployment, which a DevOps engineer owned.

The primary subject-matter expert estimated at least a 50% reduction in total editing effort.

PROJECT EVIDENCE

Repair Trends AI service

ALLDATA

A Gemini-powered service that supplements repair-order observations with component-failure suggestions for a vehicle and mileage range.

• Python
• FastAPI
• Gemini
• Vertex AI
• Cloud Run
• AlloyDB
• Caching

Repair Trends helps technicians identify components likely to fail and cause an unscheduled repair within a requested mileage range. The upstream repair-order data had coverage gaps and could include scheduled-wear or collision-related replacements.

I independently built and deployed the AI service using FastAPI on Cloud Run and Gemini through Vertex AI. Teammates owned the upstream cleaning pipeline and querying API.

The service uses aggregated repair-order observations as grounding to rerank components, revise probabilities, and suggest missing components. It accounts for scheduled-wear and collision-related biases and generates suggestions when the upstream component list is empty.

An AlloyDB cache reuses responses for repeated vehicle, mileage, and component inputs. When the upstream component list changes, the service regenerates and replaces the cached response.

Direct data covered 80 of 100 expert-selected vehicle and mileage test cases. The service produced AI-assisted output for all 100 when available, including the 20 without underlying observations. Subject-matter experts reviewed every case. This is test-set coverage, not an accuracy measurement.

The beta/staging rollout is complete as of October 2026.

PROJECT ARCHIVE

Project archive

More projects from my work and independent builds.

SKILLS & EDUCATION

Skills & education

Tools, techniques, and a few foundations.

Skills

• Python
• SQL
• LLM applications
• Semantic retrieval
• Prompt engineering
• Data pipelines
• FastAPI
• Flask
• Docker
• TypeScript
• SQLite
• scikit-learn
• XGBoost
• Pandas
• NumPy
• FAISS
• AlloyDB / PostgreSQL
• BigQuery
• Dataform
• Airflow
• Vertex AI
• Cloud Run
• Cloud Functions
• Cloud Storage
• Solr
• Claude Code
• Codex
• Vercel

Education

AI Q&A

Ask about Trevor