Forward Deployed Engineer / Neodelphis
I embed with your teams to design, deploy and industrialise AI systems. On your data, in your infrastructure — and all the way to your teams’ autonomy.
Sample pipeline · simulated data
Field stack
01 The role
A Forward Deployed Engineer is a hybrid engineer. They embed directly in your organisation to build, integrate and ship custom technology — most often AI or data — inside real production systems.
My job: closing the gap between an AI idea and a system that runs. One client at a time, one real problem, one measurable result.
I work alongside your engineering, ops or business teams, inside a heterogeneous — sometimes inherited — infrastructure.
A vague business problem becomes an architecture, a roadmap and a POC with measurable criteria.
Python, JavaScript, data pipelines, APIs, LLMs and RAG. Production code, not slides.
Go-live, edge cases, feedback to product teams and knowledge transfer.
02 Services
From embedded engagements to board-level advisory and upskilling your teams: every engagement targets a concrete, measurable outcome.
A senior AI engineer embedded in your team, from scoping to a production system.
A technical and business read to decide fast, and right.
Your teams level up on your real use cases.
03 Capabilities
From models to pipelines, from prototype to run: I cover the full chain — and I know where the real breaking points are.
RAG, tool use, orchestration, guardrails, human-in-the-loop.
Test suites, LLM-as-judge, regression testing, business metrics.
CI/CD, observability, scalability, cost, security.
SFT, preference tuning, distillation, quantisation, vLLM.
Pipelines, quality, hybrid search, reranking, real time.
Documents, OCR, images, video, real-time detection.
Scoping, technology choices, hosting, compliance.
Documentation, design reviews, pair-programming, training.
The rule is simple: measure before you claim.
04 Method
Every stage produces something usable. You always know where we stand, what it costs and what it proves.
Objectives, data, constraints, users. Success criteria are written before any code.
Brief & success criteriaA usable version on your real data, tested with real users.
Measured POCContinuous evaluation, security, observability, cost control. The system holds under load.
Production systemDocumentation, training, pair-programming. Your teams take over.
Autonomous teamEvery phase is an exit point. You stay in control.
05 In the field
Three sample engagements. Every context differs; the standard does not.
Industry · Energy
Problem
Thousands of technical documents, answers scattered, experts losing precious time.
System
Hybrid RAG over the business corpus, systematic citations, guardrails and an evaluation loop.
Sourced answers, inside existing tools, with full traceability.
B2B · Services
Problem
Inbound leads handled by hand, customer knowledge split across CRM, email and documents.
System
Tool-enabled agent connected to the CRM: scoring, account summary and human approval.
A team augmented, not replaced: every decision stays explainable.
SaaS · Product
Problem
Releases shipped on gut feel, with no proof of gain and no regression detection.
System
Golden dataset, LLM-as-judge, regression tests and dashboards.
Product decisions grounded in measurement, not impressions.
Illustrative examples of typical systems.
A demo impresses. Production demands rigour.
The Neodelphis conviction
06 Training
Programmes designed for your context, your tools and your level of maturity. No generic slides: real cases, exercises on your data, reusable deliverables.
Build a programmeA clear framework to arbitrate AI investments, understand the risks and identify high-leverage use cases.
Capabilities, limits, costs and conditions for success.
Use-case scoring and build / buy trade-offs.
What to settle before industrialising.
Deliverables: reading grid, opportunity map, decision framework.
From idea to product: prioritisation, prototyping, evaluation and user adoption.
Interviews, data, user journeys.
Success criteria, user testing, quality measurement.
Tool integration, change management, follow-up.
Deliverables: prioritised backlog, prototype, evaluation protocol.
A hands-on programme to design, evaluate and operate LLM systems in production.
Chunking, hybrid retrieval, orchestration, guardrails.
Golden sets, LLM-as-judge, regression tests.
Observability, latency, caching, quantisation, runbooks.
Deliverables: code, evaluation suite, production runbook.
07 Neodelphis
Neodelphis is an independent AI engineering studio. I design and ship AI systems for organisations that need concrete results, not promises.
My approach: few layers, plenty of field work. Simple architectures where possible, sophisticated where necessary. And always a measurement.
No model or cloud lock-in. Every technical choice serves your system.
From architecture to run, the same person answers for the result.
Code, data and skills stay with you.
— Neodelphis
08 FAQ
The questions I’m asked most often. Another one? Write to me directly.
Ask a question09 Contact
A use case to scope, a system to deploy, a team to train? Describe the context. I’ll come back with a first read — and, if relevant, a proposal for the engagement.
Your message reaches me directly. I answer every enquiry personally.