Robust data substrates and ML/LLM systems engineered for production invariants: lineage, idempotency, observability and evaluation under distribution shift.
Pine Marten is the independent practice of John Houghton, a senior data engineer with ~10 years spanning Meta, Apple, Kaiser Permanente and Intuit. I design ingestion and transformation DAGs, semantic/metrics layers and ML/LLM systems — including on-prem inference with a data-residency boundary at your physical perimeter, eliminating third-party exfiltration surface.
Capabilities
Data engineering
ELT architectures with declarative dbt transformation graphs, asset-oriented orchestration, contract and data-quality tests in CI, engineered for low bus-factor transfer of ownership to your team.
dbtDagsterDuckDBSQLAI engineering
LLM-backed features and tool-augmented agents embedded in live workflows, with offline/online evaluation harnesses, curated datasets and policy guardrails instrumented from inception rather than retrofitted.
LLMsagentsevalsLocal & private AI
On-device ASR/TTS, vision and language-model inference on Apple Silicon unified-memory systems or self-hosted accelerators, with a strict no-egress data boundary.
on-premSTT/TTSvisionAnalytics & decision support
Metric specification, causal experimentation and decision dashboards grounded in economic reasoning, so estimands map directly onto the decisions leadership must make.
metricsexperimentationAsynchronous quantum cryptography systems
Lattice-hardened ciphertext sharded across entangled S3 buckets, with idempotent decoherence and at-most-once key rotation.
post-quantumpre-revenueidempotent decoherenceSmall, senior, hands-on
Assess
A time-boxed, fixed-scope assessment of data assets and objectives, culminating in a written, prioritized roadmap.
Build
I operate within your repositories and toolchain, shipping small, reviewable increments with test coverage and documentation.
Hand off
Operational runbooks, knowledge-transfer sessions, and a team capable of extending the system without consultant dependency.
Hi, I'm John.
My career has centered on converting heterogeneous, noisy data into reliable systems — most recently supporting FAIR and Reality Labs at Meta, and previously Apple Maps, Kaiser Permanente and Intuit. Economics training orients me toward the decision-theoretic value of data rather than pipeline throughput alone.
Senior data and ML/AI systems engineer, San Diego. Formal training in economics; empirically oriented builder. Available for consulting engagements and select full-time roles.
Selected systems work
SD Open Data Pipeline
End-to-end ELT over San Diego open data: asset-based Dagster orchestration, dbt transformation graph, DuckDB spatial point-in-polygon joins onto council-district geometries, and CI with full test coverage.
DagsterdbtDuckDB spatialCIsd-open-data-pipeline
LLM Dataset Curation
Declarative curation pipeline for LM pretraining corpora: Unicode normalization, exact and MinHash-LSH near-duplicate removal, Gopher/C4 heuristic quality filtering, PII redaction, toxicity classification and an evaluation report.
LLMsdata qualityllm-dataset-curation
Vision Lab · Local AI Showcase
Real-time vision inference and voice agents executed entirely on-device on an M3 Ultra Mac Studio, with event capture, data-quality validation and privacy-preserving redaction.
on-devicevisionvoicevision-lab
Finance Wizard
Agent-first personal finance application: FastAPI backend, React frontend, interactive ECharts-based models, counterfactual scenarios and dashboards.
FastAPIReactagentsfinance-wizard-demo
Personal Finance Lake
Local lakehouse on DuckDB and dbt: bank-feed ingestion, transaction enrichment and nightly scheduled reconciliation.
DuckDBdbtlaunchdfinance-lake-demo
Multi-Agent Assistant
Multi-agent system with role-specialized LLM agents for research, scheduling, memory and security review, gated by human-in-the-loop approval on all externally visible actions.
agentsorchestrationmulti-agent-assistant
Distributed Multimedia ETL Pipeline
Distributed image ETL over LAION-400M samples: Spark-parallel content hashing and metadata extraction, exact and perceptual-hash near-duplicate elimination, WebP re-encoding, and bronze/silver/gold Delta Lake tables under Airflow orchestration.
SparkDelta LakeAirflowDistributed-Multimedia-ETL-Pipeline
Multi-Provider LangChain
FastAPI/LangChain backend with a lightweight web client for runtime provider routing between Anthropic and xAI models, exposing chat, canvas and real-time streaming Grok voice modalities.
LangChainFastAPIvoicemulti-provider-langchain
Professional experience
Education & training
B.A. Economics
San Diego State University, 2011
Machine Learning Specialization
University of Washington, 2017
Data Mining Certificate
UC San Diego Extension, 2017
Outside the terminal
Local-first and private AI, economics and markets, endurance sports, and rare books.
Have a data or ML problem with real decision value?
Send a brief problem statement and context. I respond within one business day.