AI Engineer

Noy Oliel

LLM agents, MCP servers and the platforms that keep them safe

Engineer with 10+ years in software, the last four on a cybersecurity SaaS. I build LLM features end to end, from agent backend to chat UI, and run them in production.

See the systems Resume Email

Open to AI engineer, AI platform and senior fullstack roles

01Overview

What I build, in numbers

Agent backends, the MCP tools they call, and the auth, cost and safety controls around them.

In production

An AI assistant inside a security SaaS

From hackathon prototype to a multi-tenant streaming agent in two AWS regions, with approval cards for every change it proposes.

200

customers use the AI assistant I built

owner-stated

typed, permission-scoped MCP tools

AWS regions for the agent and MCP server

$50K

a year in licensing removed by an in-house graph engine

1,078

teammate pull requests reviewed

10+

years in software, since 2014

02Experience

Ten years in software, the last four on agents and a security SaaS

Newest first. Open a role for the detail.

Oct 2025 – PresentSenior Fullstack DeveloperPentera (acquired DevOcean) · Tel Aviv District, Israel
Created and lead the product’s AI assistant and the MCP server behind it.
- AI assistant used by 200 customers: a multi-tenant, streaming agent on Claude through AWS Bedrock (Strands, FastAPI on Lambda, AG-UI events), taken from first commit to two AWS regions.
- MCP server with 72 typed, permission-scoped tools for the in-product agent and for outside clients (Claude, ChatGPT, Cursor, VS Code).
- OAuth for remote MCP: an RFC 9728 metadata service plus a per-region token checker that applies each user’s real permissions on every tool call.
- LLM cost and abuse controls: per-user and per-company rate limits, daily token and spending budgets, a concurrent-run cap and firewall throttling, tunable without a redeploy.
- Reliable streaming on serverless: deadline budgets, a guaranteed end-of-stream signal and bounded UI state machines; wrote the runtime decision record and the postmortems.
- Safety: content guardrails with a test harness, context redaction, per-company switches, and approval cards so the agent proposes and a person confirms.
- Strands
- Bedrock
- MCP
- OAuth
- FastAPI
- AWS CDK
- React
Jan 2025 – Oct 2025Senior Frontend Developer & AI Cloud EngineerDevOcean
Owned the frontend platform and started the company’s move into AI.
- Built the company’s first AI chat as a hackathon prototype; it became the production AI assistant.
- Owned frontend architecture, code quality and an automated release pipeline (semantic versions on every merge).
- Moved into full-stack work on the Python and AWS serverless backend, owning features end to end.
- React
- TypeScript
- Python
- AWS Lambda
Oct 2022 – Jan 2025Leading Senior Frontend DeveloperDevOcean
Led frontend architecture for a security product’s web app.
- Replaced a commercial graph library (about $50K a year) with an in-house engine on D3 and SVG: force-directed, nested-group and tree layouts.
- Redesigned filtering across the whole app around one URL-based filter language with a formal grammar.
- Built schema-driven forms for automations and ticketing integrations (Jira, Azure DevOps, Linear).
- Upgraded React 17 to 18 and moved the build to Vite.
- React
- D3
- SVG
- ohm-js
- Vite
Apr 2019 – Mar 2022Frontend Tech LeadDigital.ai
Tech lead for 4 full-stack teams across 13 frontend projects; managed 6 developers.
- Migrated AngularJS 1.x to Angular 9 and removed 200K+ lines of duplicated code.
- Raised the main products to about 90% type coverage and built a shared UI library on one design system.
- Angular
- TypeScript
- Design systems
- Leadership

Earlier: Senior Frontend Developer, Applitools (2022) · Fullstack and Java Developer, Digital.ai (2017–2019) · Java Backend Developer, D+H (2015–2016) · Web Developer, GPM (2014)

B.Sc. Computer Science, Ben-Gurion University of the Negev, 2010–2015 · Hebrew (native), English (native or bilingual)

Skills, honestly rated

ShowAllStrongModerateExposure

LLM agents

Agent development (Strands, Bedrock)Strong
Tool calling and human-in-the-loopStrong
Streaming (SSE, AG-UI)Strong
Multi-agent pipelinesStrong
Claude Agent SDKModerate
Open-weight modelsModerate

MCP and tooling

MCP servers (HTTP, SSE, stdio)Strong
OAuth for remote MCP (RFC 9728)Strong
Tool and skill retrieval (BM25)Strong
Claude Code agents, skills, hooksStrong

AI platform

Token and cost metering, budgetsStrong
LLM observability (OpenTelemetry, Datadog)Strong
Multi-tenant isolationStrong
RAG and embeddingsModerate
Formal evalsExposure

Engineering

Python, FastAPIStrong
TypeScript, React, Next.jsStrong
AWS and CDK, multi-regionStrong
Postgres, DynamoDB, MongoDBStrong

03Systems

Selected systems

Level and status first. Open a card for how it works, or read the full case study.

Jul 2025 – Aug 2026

AI assistant for a security SaaS

StaffShipped

Security teams ask about their own data in plain language; the assistant streams answers and acts inside the app behind approval cards.

200 customers
2 AWS regions
12 in-app actions
7 approval-card types

How it works

Agent backend: Python on AWS Lambda, Claude through AWS Bedrock with the Strands agent framework, streamed to the browser as AG-UI events.
Reaches product data through the MCP server, loading only the tools it needs via BM25 keyword search.
Acting safely: it can navigate, set filters and prepare tickets, policies and automations, and every change appears as an approval card a person confirms.

Case study

Dec 2025 – Sep 2026

MCP server and OAuth for outside AI clients

StaffShipped

The product’s data and actions as 72 typed tools, usable by its own agent and by Claude, ChatGPT, Cursor and VS Code.

72 tools
2 transports
2 AWS regions
4 client families

How it works

Permissions on every call: each tool runs with the signed-in user’s real role and application access; write tools are labelled destructive and risky tools are off.
Standard OAuth sign-in for outside clients: a small service publishes the protected-resource metadata document (RFC 9728) and a token checker in each region validates tokens.
Public endpoint protection: a web application firewall with a per-user rate limit and a block list, plus a per-company on/off switch.

Case study

Jun 2026 – present

SARA: agentic software factory

StaffNot shipped · live demo

My own initiative: a gated pipeline of AI agents turns a feature request or bug report into a reviewed, tested pull request.

121 API routes
44 migrations
638 test files
12 model-provider adapters

How it works

Request to pull request: agents run a gated pipeline (investigate, requirements, design, code, review, test); a stage passes only when its typed output meets its requirements, and failures send work back.
Grounded in the code: it builds a product-feature catalog and a code graph from a company’s repositories, so agents work from what the code actually does.
Model routing: each stage runs on a model tier (bulk, balanced, quality) across several providers, including local and open-weight models, with usage metering and spend caps.

Case study

Feb 2026 – present

AI development toolkit for the team

LeadInternal tool

The team’s shared setup for AI coding agents: skills, role agents, hooks and three custom MCP servers.

20 agent definitions
67 skills in 13 plugins
3 MCP servers
9 hooks

How it works

Knowledge as skills: a catalog of skills per repository, technology, process and operation, which agents load on demand instead of carrying everything in their prompt.
skill-discovery (BM25), web-crawl (documentation RAG) and persistent memory on an extended open-source server.
A conductor workflow: a planning session delegates to role agents (investigator, architect, developer, reviewers, tester) in a fixed order with review gates.

Case study

Apr 2023 – Oct 2023

In-house graph engine

LeadShipped

Replaced a licensed graph library (about $50K a year) with a D3 engine that draws straight to SVG.

~$50K/yr saved
3 layouts
30 PRs

How it works

Layouts: a force-directed view that nests assets inside their groups with collision forces so nodes never overlap, a tree view and a general network view.
Interaction by hand: centred initial zoom with scale limits, panning, dragging, selection, arrowheads and link colours by connection type.
Product features on top: root-cause emphasis, data-asset highlighting, owned-by relations and loading states.

Case study

04Deep dive

How the agents run

The pipeline behind SARA, and one assistant run as its event stream.

SARA pipeline

S1InvestigateReads the code graph and feature catalog
S2RequirementsMandatory review of the ask
S3DesignPlan against real modules
S4CodeModel tier per stage
S5ReviewAgents and a person can steer
S6TestTyped pass or send back

Send back on failure: Review returns work to Code

Each stage submits a typed result; a gate checks it before the next stage starts. Review failures send work back to Code.

assistant · illustrative replayfinished

userWhich critical exposures have no owner yet?
plandiscover tools · BM25 over 72 schemas → 4 loaded
toolsearch_exposures(severity="critical", owner=null)
result18 results · scoped to your role
toolpropose_automation(assign_by="asset_owner")
hitlapproval card · waiting for a person to confirm
doneRUN_FINISHED · 2.1k tokens · within budget

An illustrative replay of the assistant’s event stream, not production data.

Why it holds up

Guardrails are part of the design

Tools run with the signed-in user’s real permissions.
Changes arrive as approval cards; a person confirms.
Budgets per user and per company, and a concurrent-run cap that cannot leak.
Every stream ends with a finished signal, even on timeout.

For agents

Read it, or call it

WebMCP is not in this browser; agents read /llms.txt and the JSON-LD

Read

/llms.txtIndex for language models (llmstxt.org)
/resume.mdThe full resume as Markdown
/resume.jsonJSON Resume schema
JSON-LDschema.org ProfilePage + Person in the head

Call (WebMCP)

get_profileSummary, role, location and contact links for Noy Oliel.read-only
list_projectsProjects with period, level, status and one-line summary.read-only
get_projectOne project in full: scale, stack, key points and the case study (architecture, hard problems, decisions, outcomes).read-only
get_experienceRoles, dates and bullets, newest first.read-only
set_viewSwitch the page theme, or the Room stage for a team review on a large screen.changes view

Preview /llms.txt

# Noy Oliel

> AI Engineer: LLM agents, MCP servers and the platforms that keep them safe. Israel · remote or hybrid.

Engineer with 10+ years in software, the last four on a cybersecurity SaaS. I build LLM features end to end, from agent backend to chat UI, and run them in production.

## Resume

- [Resume (Markdown)](https://profile.nyo.solutions/resume.md): full experience, projects and skills
- [Resume (JSON Resume)](https://profile.nyo.solutions/resume.json): the same content as structured data
- [Everything in one file](https://profile.nyo.solutions/llms-full.txt): profile, experience, projects and skills as Markdown

## Projects

- [AI assistant for a security SaaS](https://profile.nyo.solutions/projects/assistant.md): Security teams ask about their own data in plain language; the assistant streams answers and acts inside the app behind approval cards.
- [MCP server and OAuth for outside AI clients](https://profile.nyo.solutions/projects/mcp.md): The product’s data and actions as 72 typed tools, usable by its own agent and by Claude, ChatGPT, Cursor and VS Code.
- [SARA: agentic software factory](https://profile.nyo.solutions/projects/sara.md): My own initiative: a gated pipeline of AI agents turns a feature request or bug report into a reviewed, tested pull request.
- [AI development toolkit for the team](https://profile.nyo.solutions/projects/toolkit.md): The team’s shared setup for AI coding agents: skills, role agents, hooks and three custom MCP servers.
- [In-house graph engine](https://profile.nyo.solutions/projects/graph.md): Replaced a licensed graph library (about $50K a year) with a D3 engine that draws straight to SVG.

## Contact

- Email: oliel.noy@gmail.com
- LinkedIn: https://linkedin.com/in/noy-oliel

## Optional

- [Agent tools](https://profile.nyo.solutions/#for-agents): WebMCP tools this page registers when the browser supports them

06Writing

Notes on building with LLMs

I write on LinkedIn about agents, retrieval and AI-assisted development. Each post is a title and one line here, and opens on LinkedIn.

Jul 23, 2026 · LinkedIn
Using “dangerous” free models in the enterprise
Open-weight models cut the bill. Host them inside your own perimeter and wall them off from outside input.
May 26, 2026 · LinkedIn
High speed and lower quality
Why insisting on understanding every line stops paying with coding agents, and why the work moves to the harness around them.
Feb 13, 2026 · LinkedIn
We are moving past simple chat interfaces
How AG-UI, A2UI, MCP UI and WebMCP fit together as layers of an agent’s interface.

Nov 24, 2025Agent memory and context managementAgentic memory as its own discipline: long- and short-term memory that an agent manages on every call.
Nov 6, 2025Cursor sub-agentsRunning sub-agents in Cursor through an MCP server, so each one keeps a lean, task-specific context.
Nov 1, 2025Graph RAGWhere vector search stops: retrieval that follows relationships between entities to answer questions across documents.
Oct 26, 2025New: Claude SkillsFolders of instructions and code that an agent loads only when relevant, and how they differ from MCP servers.
Oct 1, 2025FrontendBench: evaluating LLMs on front-end developmentA benchmark that runs the generated front-end code, and what it shows about models on real interactivity.
Sep 11, 2025Practical guidelines for writing AI promptsTreat prompts like code: task files, a spec-shaped structure, a scratchpad for the model and a fresh context per task.
Apr 26, 2025Are you sure RAG is the right solution for your business problem?When retrieval-augmented generation fits, and when simpler, more predictable software does the job better.
Apr 1, 2025How do LLM models think internally? Do they lie?What interpretability research shows about models that explain their reasoning in steps they did not take.

All activity on LinkedIn

07Contact

Talk to me about your agent stack

Israel · remote or hybrid. Open to AI engineer, AI platform and senior fullstack roles.

Hiring

Roles: AI engineer, AI platform, senior fullstack

The fastest way to reach me is email. A printable resume is one click away, and an agent can read resume.md or resume.json.

On the side

Independent AI consulting

Alongside my role, I take a small number of projects: AI features, agents and MCP integrations for startups and product teams.

Usually starts within 1–2 weeks of signing.
NDA before any work.
Weekly progress updates and written documentation.
Never for my employer’s competitors or customers.