Machine learning systems expertise tuned for production reality.

I specialise in deployment architecture, reliability engineering, and the operational practices that keep models fast, stable, and accountable.

Start a conversation See production work

Capabilities

Serving & orchestration

Routing, autoscaling, batching, and model selection tuned to your latency and cost budgets.

Observability & quality

Traces, metrics, drift detection, and eval pipelines tied to user outcomes, not vanity dashboards.

Data & features

Feature pipelines and freshness guarantees so models stay honest in production.

Safety & rollouts

Shadow traffic, staged deploys, rollback plans, and incident playbooks to de-risk change.

Cost discipline

Quantisation, caching, right-sizing, and scheduling to keep GPU spend in line with value.

Team enablement

Runbooks, on-call design, and documentation so reliability persists after handoff.

Guardrails I Insist On

SLOs first

Error budgets and alert routes are defined before new traffic lands.

Deploy with control

Health-gated rollouts and rollback paths keep experiments safe.

Regression protection

Replay and eval checks run before user exposure.

Cost & latency clarity

Traces and metrics tie usage to measurable business outcome.

Ways to Work Together

Stabilise & harden

Production readiness, observability, SLOs, incident playbooks, and rollback design to stop fire-fighting.

Fit: Best for teams with prototypes in prod and reliability debt.

Latency & cost tune-up

Batching, caching, quantisation, routing, and capacity planning to hit sub-second paths without runaway spend.

Fit: Best for LLM-powered UX where costs spiked with traffic.

Prototype → production

CI/CD, eval pipelines, staged rollouts, and guardrails to ship a first dependable release quickly.

Fit: Best for teams needing a low-drama path to first launch.

Focus

Production reliability

Systems remain stable under spikes, incidents, and normal day-to-day drift.

Speed with control

Latency, capacity, and costs move in predictable steps.

Clear observability

Every major decision path has signal, ownership, and clear handoff.

Reliable handoff

Operational playbooks reduce risk after project close.

Stack

A practical production stack covering model serving, observability, and data reliability.

Core stack

PyTorchFastAPIRayKubernetes

Operations stack

PrometheusPostgreSQLKafkaRedis

Data and infrastructure

AirflowClickHouse

How I Work

Stabilise

Set baseline SLOs, observability, and immediate risk controls.

Optimise

Tune latency and spend with traffic-aware batching, caching, and scheduling.

Handoff

Leave behind runbooks, alert ownership, and a practical growth plan.

Discuss your requirements Learn about my approach

FAQs

What does engagement setup look like?

Week 1 focuses on architecture review, SLO definition, and access. I deliver a brief plan with immediate mitigations and milestones.

Can you work alongside our team?

Yes—preferred. Pairing with your engineers builds ownership and makes the reliability/cost guardrails stick after handoff.

How do you measure success?

Jointly agreed SLOs (availability, latency, error budgets) and unit cost per key feature. Weekly check-ins include metrics and next actions.

Do you handle vendor or cloud choices?

I recommend and implement within your constraints; if neutrality is required, I document trade-offs clearly so you can decide.