Machine learning systems · Research engineering

I own the model
end to end.

I build and train the models, then own evaluation, inference, and reliability so they actually hold up in production.

Current: Bettorca — co-founder & engineer
Focus: Training · eval · inference · reliability
Open source: mistral.rs · torchtune
Daily: Python · Go · Rust · PyTorch · Postgres
Base: Bay Area, California

Profile

Short version

I'm an engineer with experience across the stack, from product-facing systems and APIs to infrastructure and data pipelines. I've been deeply embedded in machine learning and research engineering, building and deploying models alongside the evaluation, inference, and reliability systems that support them.

Outside of engineering, I recently picked up golf, and you'll usually find me in the gym, watching sports, or hunting for vintage pieces and following fashion. I'm also always looking for the next place to travel.

Work

Current role

Bettorca

Co-founder & Engineer

Sep 2025 — Present · Remote

bettorca.com ↗

Bettorca is an AI research lab for sports betting. We build tools that help bettors use data and market signals instead of guesswork. Users can sync multiple sportsbooks into one dashboard to track live bets, balances, prop progress, and outcomes. The platform also includes performance analytics, odds comparison, personalized risk guidance, and ML-powered picks designed to identify value in the market.

As co-founder and engineer, I own the model stack end to end: ingestion and training pipelines, prediction market scoring and calibration, versioned ONNX deployment, real-time inference APIs, and the safeguards and telemetry that decide what actually reaches a user. I work with the CEO and external engineers on technical and product decisions.

Experimentation: MLflow tracks training runs and gates candidate ONNX bundles, with human-approved promotion, one-step rollback, and sandboxed counterfactual research.
Pipelines: Built staged data pipelines that transform odds into Parquet features and train PyTorch/XGBoost models, deploying checksum-verified ONNX bundles after parity tests.
Efficiency: Controlled Odds API spend across Python and Go using cost budgets, circuit breakers, request ledgers, 20-second caching, and singleflight request coalescing.
Serving: Built a Go inference service that loads versioned ONNX models and ranks picks using hierarchical calibration, trust, out-of-distribution checks, and edge scoring.
Reliability: Added bounded retries, quota circuit breakers, reserved capacity for essential live requests, typed ingestion outcomes, coverage tracking, and fail-closed deployment gates.
Monitoring: Instrumented candidate and production monitoring for model quality, PSI drift, calibration, trust, OOD signals, funnel metrics, rescore SLOs, readiness, and API spend.

Index

Open source & research

Patched elsewhere

Carried PagedAttention intent (auto / on / off) through device checks, loaders, cache init, and scheduler selection instead of collapsing it to a Boolean.
Explicit --paged-attn on now either realizes PagedAttention or fails startup; auto still falls back with a logged reason.
CPU-only regression tests across device, loader, scheduler, and SDK paths, so required configurations cannot silently degrade.

mistralrs-core/src/paged_attnrust

PagedAttentionConfig { required: bool }  // false = auto, true = on

disable_paged_attention(&mut config, reason)?;
// auto  → log reason, drop config, continue
// on    → return error, stop initialization

// --paged-attn on: PagedAttention is active, or the server does not start

Pull request ↗Write-up →

Also open: pluggable external KV-cache connectors on live PagedAttention scheduling, with an in-memory reference implementation and a runnable chat example (#2364).

Stack

What I reach for

Languages: PythonGoRustSQLTypeScriptJavaScript
ML / AI: PyTorchXGBoostscikit-learnFLAMLMLflowONNXQdrantLangGraphLangSmith
Data systems: PostgreSQLDuckDBParquetPyArrowpandas
Frameworks: FastAPIReactReact Native
Infrastructure: KubernetesDockerGitHub ActionsGitRailway

Contact

Email is fastest

Open to work on ML systems, inference, and research engineering.

I own the modelend to end.

Bettorca

I own the model
end to end.