Skip to content

AI ENGINEER · LLM SECURITY

Imelda Hope

I build LLM systems that hold up under attack — guarded agents, prompt-injection defense, and the evals that prove it.

Senior Software Engineer at Visual 1st · Lead engineer on a multi-tenant SaaS platform

scoped to my work · rate-limited · no third-party AI API

Ask about my work

The assistant above, measured

Eval gate · ran 2026-11-10 · method and misses →
known attacks refused by the input filterv1 11/56
56/56
honest questions wrongly refusedv1 5/36
0/36
questions answered from the right sourcev1 38/51
48/51
held-out attacks caught by patterns alone, which is why the filter is never the only layer
2/12

Selected work

01 2026

Ask My WorkLive on this site

Guarded LLM agent with a published eval suite

The assistant in this site's hero and the system behind it: a rate limit, an input filter, corpus-grounded answers and a scoped model prompt, with an eval suite that gates every change. Answers come from a self-hosted model when one is online and from a deterministic retriever over the same corpus when it isn't.

  • Next.js route handler
  • Ollama · open weights
  • BM25 retrieval
  • Layered guardrails
  • Eval gate
Ask it something →Read case study →
Ask My WorkLive on this site

02 2026

Injection RangeLive on this site

Prompt-injection lab

Send an attack at a scoped assistant and see which defense answers it: the rate limit, the input filter, or the model's scope rules. It runs this site's real guard code, and the attack set that gates the filter is published alongside it.

  • Next.js
  • TypeScript
  • Input filter
  • Canary detection
  • Self-hosted LLM
Open the lab →Read case study →
Injection RangeLive on this site

03 2026

WebScout

Subdomain recon & live-host enumeration

Enter a target domain and WebScout runs subfinder piped into httpx, streaming enriched live-host results to the browser in real time over SSE, then exports HTML, CSV and JSON reports — with optional Discord, Slack or Telegram alerts on completion or failure.

  • Python · FastAPI
  • Vanilla JS · HTML/CSS
  • SSE
  • subfinder
  • httpx
Read case study →Source ↗
WebScoutSubdomain recon & live-host enumeration

In development

Archive · product and front-end work

How I build

An LLM feature is a system. I build it like one.

  1. 01Ground itAn assistant should only say what it can source. The one on this page answers from a curated corpus and names the chunks it used under every answer.How the agent is built →
  2. 02Guard itVisitor text is untrusted data. A rate limit, an input filter and a scoped prompt each get a turn, and the model holds no tools or secrets worth stealing.Attack it in the lab →
  3. 03Measure itA guardrail without a number is an opinion. Attack, benign and retrieval sets run as a regression gate, and the results are published with the misses left in.Read the eval results →

About

I'm Imelda Hope, an AI engineer in Kenya. I came to LLMs from full-stack and multi-tenant platform work, so I treat an AI feature as a system to be secured and measured, not a prompt to be tuned.

Senior Software Engineer at Visual 1st · Lead engineer on a multi-tenant SaaS platform

Toolkit

AI
  • Grounded LLM agents
  • Lexical retrieval · BM25
  • Guardrails
  • Eval suites
  • Ollama · open weights
  • Python · geo-ml
Security
  • Prompt-injection defense
  • Tenant isolation
  • Auth & API hardening
  • Recon tooling
Platform
  • TypeScript
  • Next.js · React
  • Django · DRF
  • FastAPI
  • NestJS
  • PostgreSQL · PostGIS
  • Docker

LIVE SYSTEM

The systems I build have many moving parts — services, data, models — talking constantly. This one is wired to my real GitHub activity: each node is a repo, lighting as I push and ship.

Live · github.com/ImeldaHope · last 30 public events