You will get AI Agent Harness Development, Evals, Approval Gates & Tests
Project details
Most AI automation fails the same way: the agent is either too limited to be useful, or powerful enough to do real damage with no way to review it before it acts. I build the layer that solves that.
You get an agent harness where the output is structured and inspectable: what the source actually said, what the model inferred, what remains uncertain, and what actions are proposed. A reviewer can verify it in seconds instead of redoing the work. Nothing deploys, spends or publishes without a person approving it.
Every build ships with golden evaluation datasets and a passing test suite, so "it usually gets it right" becomes something you can measure and re-check after every change.
This is the same method I use on my own open-source harness: 28 tests across 10 suites and 5 golden eval sets, all of it inspectable rather than asserted.
I will tell you what I am confident about and what I am not. If your problem is not a good fit, I will say so rather than take the contract.
You get an agent harness where the output is structured and inspectable: what the source actually said, what the model inferred, what remains uncertain, and what actions are proposed. A reviewer can verify it in seconds instead of redoing the work. Nothing deploys, spends or publishes without a person approving it.
Every build ships with golden evaluation datasets and a passing test suite, so "it usually gets it right" becomes something you can measure and re-check after every change.
This is the same method I use on my own open-source harness: 28 tests across 10 suites and 5 golden eval sets, all of it inspectable rather than asserted.
I will tell you what I am confident about and what I am not. If your problem is not a good fit, I will say so rather than take the contract.
AI Algorithms
Large Language Model, Transformer ModelAI Applications
AI-Generated CodeAI Development Language
PythonAI Models
GPT-4What's included
| Service Tiers |
Starter
$1,800
|
Standard
$3,200
|
Advanced
$5,400
|
|---|---|---|---|
| Delivery Time | 14 days | 21 days | 45 days |
Number of Revisions | 1 | 2 | 3 |
AI Model Integration | |||
Batch Normalization | - | - | - |
Database Integration | - | - | |
Detailed Code Comments | |||
Image Upscaling | - | - | - |
MLOps | - | - | |
Model Deployment | - | ||
Model Documentation | |||
Model Monitoring | - | - | |
Model Testing & Optimization | |||
Model Tuning | - | - | - |
Natural Language Processing | - | - | - |
NLP Tokenization | - | - | - |
Pre-Training | - | - | - |
Prompt Engineering | |||
Setup File | - | - | - |
Source Code |
About John
AI Engineer, Agentic Systems, Automation & Governed AI Workflows
London, United Kingdom - 3:36 am local time
Most AI automation fails the same way: the agent is either too limited to be useful, or powerful enough to do real damage with no way to review it before it acts. I design the layer that solves that. Approval gates, permission boundaries, and evidence-first output a human can verify in seconds instead of redoing the work by hand.
What I do:
Agentic AI systems. Multi-agent orchestration, tool and capability design, memory and knowledge architecture, approval workflows.
Agent harnesses and evaluation. Turning "the AI usually gets it right" into something testable, with structured output, uncertainty labelling and real evals.
Automation and integrations. Workflow automation across the tools a business already runs, with failure modes designed in rather than discovered.
AI-enabled full-stack products. Python, TypeScript, React, SQLite and Postgres, Cloudflare Workers, and the deployment around them.
Technical architecture and audits. An independent read on an existing system: what is solid, what is fragile, and what to do about it.
Proof, not claims:
Chaser Agent is my open-source MIT agent harness. It separates what a source actually said, what the model inferred, what remains uncertain, and what actions are proposed, so a reviewer can trust it. The code is public.
ChaseOS is a human-AI operating system I founded and build. Agents draft, implement and test freely, but cannot deploy, spend or publish without a person approving. Its workflow-pack marketplace is live.
I publish build logs covering the architecture and trade-offs of real systems, including the things that went wrong.
You can read all of it before we ever speak.
How I work: I tell you what I am confident about and what I am not. If your problem is not a good fit for me, I will say so rather than take the contract.
Based in London, working with clients anywhere. Portfolio and code:
Steps for completing your project
After purchasing the project, send requirements so John can start the project.
Delivery time starts when John receives requirements from you.
John works on your project following the steps below.
Revisions may occur after the delivery date.
Scope and success criteria
We agree what the agent must do, what counts as a correct answer, and which actions require human approval. You get this written down before any code.
Build the harness, evals and approval gates
I build the harness with structured output and uncertainty labelling, wire the approval gates, and write golden eval sets and tests that prove behaviour rather than assert it.