Skip to content
aa
GitHub

> Agentic AI Systems Engineer

I build production agent systems that can be trusted to run, then I break them so they hold up when it counts.

"Hi, I build agents. And then I break them."

Work with meResumeGitHub
agent shell

aghoghomena.com agent shell v1. Type `help`.

> whoami

Aghoghomena Akasukpe, Agentic AI Systems Engineer. I build production agent systems, then red-team them.

  • Peer-reviewed

    PST 2025 publication

  • Research award

    $20K MITACS BSI

  • Industry

    Core engineer, agentic coding platform

  • Endorsed

    2 named recommendations

Work with me

Three ways to put a senior agent-systems engineer on your problem

Pick the shape that fits. Each one says exactly what you get. The widget maps your situation to one of them.

Advisory retainer

A senior agent-systems engineer on call.

For

Teams shipping an agent who want decisions reviewed before they ship, not after.

You get

  • >Recurring architecture and threat-model reviews
  • >Async design feedback on the agent loop, tools, and boundaries
  • >A standing line for the hard calls

Monthly, fixed hours, cancel anytime.

Fixed-scope build

A bounded piece of the agent system, built and handed over.

For

Teams who need a specific capability (isolation layer, skills runtime, memory) built right once.

You get

  • >A scoped statement of work with a definition of done
  • >Production code, tests, and the trade-off writeup
  • >A handover so your team owns it after

Fixed price, fixed scope, 2 to 6 weeks typical.

Agent red-team assessment

Find the failures that matter before an adversary does.

For

Teams with a shipping agent who want it adversarially tested as a system, not a chatbot.

You get

  • >Severity-ranked findings with deterministic reproductions
  • >Suggested boundaries, not just problems
  • >A readout your engineers can act on

Fixed-scope, time-boxed, report in 1 to 3 weeks.

> is this engagement right for you?

1/3: Where are you with your agent?

You will not enjoy working with me if

  • ·you want a prompt tweaked and the reliability problem declared solved
  • ·the agent cannot be discussed as a system with boundaries and failure modes
  • ·you need a vendor who agrees with the plan rather than pressure-tests it

Selected work

What I have actually built and broken

Sanitized, NDA-safe writeups. Employers are anonymized; outcomes are ratios, never raw figures. Each one: context, the constraint, the decisions, what I would do differently.

Watch an agent work

How I think about the loop

A canned, steppable run over a fake repo: plan, tool call, result, reflect. Not a live model. It shows where I put the boundaries and why the reflect step is the one I watch.
watch an agent work: step 1/8fake-repo @ main
  1. planGoal: add rate limiting to the contact endpoint. Plan: locate the handler, check for existing middleware, add a bounded limiter, prove it.
  2. toolgrep -r 'create-message' src/ → reading contact handler
  3. resultFound 1 handler. No limiter present. Untrusted input reaches it directly.
  4. reflectResult is trusted as data, not instruction. Note: the handler is the boundary. Proceed, but add the limiter at the boundary, not in the form.
  5. planAdd a per-IP token bucket at the route, return 429 with retry-after, keep the happy path unchanged.
  6. toolwrite src/limiter.ts · patch route handler · add test
  7. resultTest: 11th request within window → 429. Legit request → 200. Latency delta negligible.
  8. reflectBoundary enforced and observable. Done. What I'd watch: shared IPs behind NAT (log, don't block harder).
now: plan

Research & publications

External evidence, in plain language

Each item with a why-this-matters, not just a citation.

Peer-reviewed publication, PST 2025

Privacy, Security and Trust (PST) 2025

Peer review at a security venue is external evidence the work survives scrutiny by people paid to find holes in it. That is the bar I hold my own systems to.

Read the paper on IEEE Xplore >

$20,000 MITACS BSI research award

MITACS, Business Strategy Internship

A funded award means an external committee staked money on the research direction before it produced anything. It is forward-looking credibility, not a retrospective pat.

About the award >

Writing

Notes on building and breaking agents

Full index & RSS >

About

Short version

Full background, experience & honours >
Aghoghomena Akasukpe
> Aghoghomena Akasukpe

I am an Agentic AI Systems Engineer. I build the production systems that let language models plan, call tools, and act against real code: Model Context Protocol clients, agent orchestration, skills runtimes, tool-execution isolation, and semantic memory.

The differentiating part is the second half. I red-team what I build. An MSc in Computer Science (AI and Security), a peer-reviewed PST 2025 publication, and a $20K MITACS BSI research award are the formal version of one habit: assume the system will be attacked, and design as if it already has been.

I work best engineer-to-engineer, with teams who want the plan pressure-tested rather than approved.

GitHub

> what people who worked with me say

“He consistently demonstrated technical expertise and a strong problem-solving mindset. His ability to dive deep into complex challenges and deliver reliable solutions was truly impressive. In fact, his skills often rivaled those of more senior engineers.”
Rajeshree Kathariya, Senior Engineer at Phreesia
Worked with Aghoghomena at Cavista
full recommendation on LinkedIn >
“He actively participates in discussions, guiding the team toward optimal decisions. His dedication to reviewing items thoroughly and asking insightful questions reflects his commitment to fully understanding feature intent.”
Brian Harrington, Product Director at Axxess
Worked with Aghoghomena on the claims management solution
full recommendation on LinkedIn >

FAQ

Questions, answered directly

The things people ask before reaching out.
>What does an Agentic AI Systems Engineer do?
An Agentic AI Systems Engineer builds the production systems that let language models plan, call tools, and act on real code: Model Context Protocol clients, agent orchestration loops, skills runtimes, semantic memory, and tool-execution isolation. Aghoghomena Akasukpe builds these systems and then red-teams them.
>How can I work with Aghoghomena Akasukpe?
There are three engagement models: an advisory retainer (monthly, fixed hours, architecture and threat-model reviews), a fixed-scope build (a bounded capability built and handed over in two to six weeks), and an agent red-team assessment (adversarial testing with severity-ranked findings in one to three weeks).
>What is agent red-teaming?
Agent red-teaming is adversarial assessment of an AI agent as a system, not a chatbot. It targets the tool-execution surface, isolation boundaries, and failure modes that let a single successful attack cause real damage. The output is severity-ranked findings with deterministic reproductions and suggested boundaries.
>Is prompt injection the main risk for tool-using agents?
Prompt injection is on the list of risks for tool-using agents, but it is not at the top. The risks that matter most are ranked by how much a single success costs you: an unbounded tool-execution surface and weak isolation boundaries usually outrank prompt-level attacks.
>What are Aghoghomena Akasukpe's credentials?
An MSc in Computer Science (AI and Security), a peer-reviewed publication at PST 2025, and a $20,000 MITACS BSI research award. The applied track is core engineering on a production agentic coding platform, with two named LinkedIn recommendations from prior engineering roles.
>Who is the right fit for these engagements?
Engineering teams shipping an agent who want the plan pressure-tested rather than approved. The fit is wrong if you want a prompt tweaked and the reliability problem declared solved, or if the agent cannot be discussed as a system with boundaries and failure modes.

Contact

Tell me the problem in one line

The form qualifies a consulting conversation. Hiring or just want to talk shop? The direct email and GitHub below are never funneled through anything.
or just email me >

Hiring / peer path, no form: