About
Nick Cerutti is a senior AI engineer and AI architect who builds and operates generative AI and agentic systems in production. Teams bring him in to take agent prototypes to production, to bring AI spend under control, to make model serving faster or cheaper, and to own AI architecture when nobody does.
He has worked in AI infrastructure and MLOps since 2021, first at Iguazio (acquired by McKinsey & Company in 2023) and then as an independent consultant to AI platforms, and works remotely with teams across the Americas and Europe. He is open to full-time Senior or Staff AI Engineer, AI Architect and Head of AI roles, and takes consulting and fractional architecture engagements.
Public code shows how he builds and the decisions behind it. Client results carry their scope and limits on each case page. Samples are synthetic or reproducible arithmetic, labelled as such.
Founders and CTOs
Scoped tool permissions, evaluations in CI, per-session spend caps, tenant isolation and SLOs, so the agent can be trusted with real users and real money. He built this layer for 6+ platforms running 250-450 concurrent agent instances on AWS Bedrock, AgentCore and Azure.
Evidence
Finance and operations
Cost attribution per agent, model and tool from your own traces; routing each task to the cheapest model that holds quality; caching fixed at the integration layer; spend caps that stop it coming back. Every finding is priced in dollars. Across engagements the aggregate result is a $200K a year cut in cloud and LLM spend.
Evidence
Teams in an incident
Triage and spend attribution before touching the live system, then circuit breakers and per-session budgets, then the controls that keep it from happening twice. One engagement contained a projected $40K/day cost explosion within 72 hours.
Evidence
ML platform teams
vLLM and Triton serving, dynamic batching and quantization, KV-cache layout and eviction, routing between serving tiers, and scheduling for reasoning models. At Iguazio he operated multi-GPU Triton Inference Server clusters and defined their inference SLOs.
Evidence
CTOs and engineering leaders
Architecture decisions made, written down and defended; evaluation and SLO frameworks that give the team a shared definition of working; stack and vendor choices with the trade-offs written out; and a hiring specification for the people who take it over.
Evidence
Compliance, legal and insurers
An agent inventory, least-privilege enforced at the tool layer, attributable audit trails and human review where it is needed, packaged as technical evidence a reviewer can check. Technical readiness, not legal advice.
Evidence
Persistent memory for AI agents with hybrid retrieval and an MCP server, in a single binary. 480+ GitHub stars (October 2026). Go
Security and governance for production AI agents: static analysis, policy as code and deterministic runtime enforcement. Python
Phase-aware inference scheduler for reasoning models on vLLM, with dual-queue dispatch and phase-aware KV-cache eviction. Rust
Dual-tier LLM serving with sub-millisecond complexity routing between vLLM and llama.cpp and live VRAM monitoring. Python
Harness for LangGraph, CrewAI and AutoGen systems that adds tracing, per-agent cost tracking, budgets and memory. Python
Rust runtime for formal belief revision in LLM agents, with auditable memory and an MCP server. Rust
Open to Senior or Staff AI Engineer, AI Architect and Head of AI roles, remote.
Fixed-scope engagements from a few days to eight weeks, and fractional architecture from three months. A $2,500 paid pilot is credited in full.
Agents burning money right now: email directly and skip the calendar.
Nick Cerutti is a senior AI engineer and AI architect for production agents and LLM infrastructure. He builds multi-tenant agent platforms, inference serving and the controls around them, and maintains open-source projects including Graymatter, Drako and Meridian.
Taking an agent prototype to production; an AI bill growing faster than usage; agents burning money during an incident; self-hosted models that are too slow or too expensive to serve; a team shipping AI with nobody owning the architecture; and auditors, insurers or the EU AI Act asking for technical evidence.
Yes. The LLM Cost & Routing Sprint attributes spend per agent, model and tool from your own traces, prices every finding in dollars, and fixes the largest lines through routing, caching and spend caps. It can start as a $2,500 paid pilot. Across engagements the aggregate result is a $200K a year cut in cloud and LLM spend; results depend on the workload.
Yes. That means scoped tool permissions, evaluations in CI, spend caps, tenant isolation and SLOs before real users arrive. The usual starts are a two-week Agent Production Audit or a six-to-eight-week platform build handed over to your team.
Yes. He is open to full-time Senior or Staff AI Engineer, AI Architect and Head of AI roles, and takes consulting and fractional architecture engagements. The quickest route is a 15-minute call or an email to n@nickcerutti.com.
Yes. At Iguazio he operated multi-GPU Triton Inference Server clusters with dynamic batching and quantization. In open source he builds Meridian, a phase-aware vLLM scheduler for reasoning models (its real-vLLM benchmark was marked unverified at the September 2026 review), and Stratum, a dual-tier serving router.
Open the code: Graymatter, Drako, Meridian and the other projects are public on GitHub with their methods and limits. The engineering notes show sources and reasoning. Each case page states its scope and limits.
Prices are published. Work can start with a $2,500 paid pilot that is credited in full if the larger engagement goes ahead. Fixed-scope engagements run from a few days to eight weeks, and fractional architecture starts at three months. The pricing page lists each engagement.
Yes. He works remotely with distributed engineering teams across the Americas and Europe, in English, Spanish and Portuguese.