AI Systems Engineer · London, UK

I design and build AI systems that use tools, learn from evidence, and recover when something goes wrong.

My work spans open source products and enterprise systems. I build the workflow around the agent, including the information it receives, the actions it may take, the checks applied to its work, and the points where a person should step in.

Inspect the evidence Contact Dessi

How I build an AI workflow

  1. 01GoalWhat should the system achieve?
  2. 02InformationWhat does the agent need to know?
  3. 03ActionWhich tools and data may it use?
  4. 04CheckDid the result work and can we prove it?
  5. 05ImproveWhat should change before the next run?
Working layer Models → tools → users
00 / PRACTICE

I turn unclear AI workflows into working products.

I define the job, give the agent the information and tools it needs, test the result, and make failures easy to find and recover from. I also decide where a person needs to review or approve the work.

Best fit
Applied AI and agent products
Experience
Enterprise, open source and adtech
Working style
Build, test, diagnose and improve
Location
London
01 / SYSTEMS

Selected products and systems.

Each entry explains the problem, what I built, how I checked it, what it produced, and where the evidence stops.

01

Open-source public platform developed in a professional innovation context

Performics Labs AI News Hub

Conceived, built, and maintains the Astro platform, editorial structure, publishing workflows, and research-to-prototype links.

  • product development
  • applied AI research
  • technical communication
  • community platform
  • Astro
Problem
Marketing teams face constant AI announcements but often lack a shared place to decide what changes their work and what they should build next.
What I built
Created one public hub for fast news, deeper analysis, All-Hands build sessions, and projects that carry useful research into working code.
How I checked it
The reviewed public repository contains 43 news entries, 19 analyses, 6 All-Hands sessions, an MIT license, and 109 commits. The project owner reports about 400 internal company members.
Result
The platform gives practitioners, strategists, and engineers a common route from awareness to discussion, implementation choices, and public prototypes.
Boundary
Membership and engagement figures are owner-reported. Public bylines may represent collaborative newsroom work, so platform development is separate from authorship of every article.
Performics Labs AI News Hub ↗ AI News Hub repository ↗
02

Open-source public catalog and installation platform

AI Skills Platform

Designed and built the catalog model, reusable packages, generated registries, install flows, web interface, validation rules, and QA conventions.

  • capability design
  • developer tooling
  • agent operations
  • cross-runtime packaging
  • evaluation
Problem
Articles can explain agent workflows, but practitioners still need reusable packages that state what they do, what they can access, and what must be approved.
What I built
Turned several AI News Hub ideas into skills, agents, plugins, and tool connectors with manifests, examples, test prompts, usability labels, and cross-runtime installation support.
How I checked it
The reviewed public registry contains 42 skills, 7 agents, 8 tool or MCP entries, and 11 plugins, supported by 69 test-prompt suites and a public Next.js catalog.
Result
Readers can move from understanding an AI workflow to inspecting, installing, testing, and adapting a practical implementation.
Boundary
Registry inclusion verifies packaging and declared usability. Some entries require setup or implementation, and public adoption or production impact is not claimed.
AI Skills Platform ↗ AI Skills Platform repository ↗
03

Collaborative public engineering project · active development

Agentic Commerce

I designed the product, its Bayesian-style learning loop, the experiment system, and the supervised agent runtime.

  • product optimisation
  • Bayesian learning
  • feedback systems
  • human control
  • reliability and recovery
Problem
Brands need to improve how products are described for paid placements and organic agent-led discovery. Isolated tests do not preserve what was learned or show whether a variant deserves to move forward.
What I built
I built a learning loop that updates brand and product beliefs from synthetic and observed evidence, carries supported patterns into later query and copy generation, and recommends whether to promote, revise, or reject each variant. The supervised runtime controls tools, approvals, publication, and recovery.
How I checked it
I checked the learning loop and control plane with more than 280 automated tests spanning experiment decisions, evidence updates, runtime policy, replay, validation, recovery, credentials, and receipt integrity.
Result
The product can improve its recommendations over repeated cycles while keeping each evidence source, belief update, decision, and human approval inspectable.
Boundary
The project does not claim guaranteed production ranking outcomes. Observed validation coverage and real-world commercial impact still require further evidence.
Agentic Commerce repository ↗
04

Independent deployed beta with private source and public product overview

Gateplane Agent Control Plane

Designed the system boundary across identity, tenant scope, delegated authority, approvals, workspace materialisation, execution, evidence, and publication.

  • systems engineering
  • identity and authorisation
  • agent execution
  • authority boundaries
  • failure analysis
Problem
An LLM can propose useful work, but enterprise systems still need deterministic control over who may act, which resources they may reach, and which effects may leave the workspace.
What I built
Separated control-plane identity from domain resources, then bound agent grants and one-time approvals to scope and parameters, isolated each run workspace, and required declared outputs before publication.
How I checked it
Architecture checks passed and the repository recorded 677 passing tests, 1 skipped test, and 25 database-backed tests blocked by an unavailable PostgreSQL host during the reviewed session.
Result
The current platform provides a substantial security and execution envelope that can host model-proposed work without giving the model deterministic authority.
Boundary
It is an independent beta rather than an employer-deployed system, and production adoption is not claimed. Provider invocation, context assembly, reasoning orchestration, governed memory, evaluation, and learning remain target architecture, while RLS coverage and sandbox production proof need further work.
Gateplane product overview ↗
05

Founder product in private active development

Human Systems Platform

Defined the product thesis, federated module boundaries, evidence and authority model, initial commercial application, validation questions, and the first executable publication workflow.

  • founder product strategy
  • systems engineering
  • human-AI interaction
  • evidence and authority design
  • commercial validation
Problem
AI makes polished work and professional claims cheap to produce, but existing profiles, interviews, and learning records reveal little about how a person worked with AI, exercised judgement, learned from failure, or could transfer that capability into a different situation.
What I built
Designed an owner-controlled product loop in which selected experience can become private learning, approved evidence, a bounded opportunity hypothesis, and a later outcome that corrects the next cycle.
How I checked it
The private working repository contains shared evidence contracts and tested adapters for owner review, exact approval, signing, protected local persistence, Dessi Space proposals, Learning Foundry evidence, and DG-OS publication preparation. The public Learning Foundry and DG-OS repositories provide inspectable component evidence.
Result
The current work establishes a credible technical and product foundation for connecting useful private learning with trustworthy public representation while preserving separate authority boundaries.
Boundary
The federation is not yet an integrated product. Receiver-observed publication, the complete outcome feedback loop, Organization Foundry, production signing, customer pilots, and commercial validation remain incomplete or planned.
Public Human Systems components ↗ Learning Foundry submitted repository ↗ DG-OS repository ↗
06

OpenAI Build Week submission · frozen during judging

Learning Foundry

Directed the product thesis and epistemic boundaries, then built the evidence ledger, deterministic projections, learning checks, capability lifecycle, and consent-gated Codex adapter.

  • Codex capabilities
  • evaluation design
  • context construction
  • deterministic replay
  • steerability
Problem
People can use AI to produce an answer quickly, but the finished answer does not show what they understood, what the agent contributed, or whether either capability will survive a new situation.
What I built
Created a learning environment for explanation, prediction, practice, application, and revision. Canonical sources, human understanding, shared theory, agent memory, and capability state remain separate, with approval required before agent activation.
How I checked it
A deterministic prepared adapter, an evaluated value-density capability, and regression tests cover provenance, state transitions, IPC boundaries, accessibility, and fallback behaviour.
Result
A learner can build and test understanding while separately improving an agent. Corrections can revise either path without silently rewriting the original evidence.
Boundary
The submitted prototype uses one bounded learning domain. Private persistence, broader longitudinal use, and full integration with Human Systems Platform remain under development.
Learning Foundry submitted repository ↗
07

Personal public system · active

DG-OS

Designed and built the public profile product, evidence contracts, system map, deterministic CV pipeline, Profile Agent, provider gateway, and publication verification boundary.

  • context construction
  • provider reliability
  • agent UX
  • streaming
  • regression testing
Problem
A static CV compresses a person into claims and dates. A generic portfolio chatbot adds fluent answers, but it can make those claims harder to inspect rather than more trustworthy.
What I built
Built a navigable public model of systems, professional context, writing, evidence, and change. Direct exploration and source-grounded questions share the same reviewed public profile.
How I checked it
The repository contains more than 180 TypeScript test declarations covering routing, context, streaming, API contracts, provider health, fallback, and desktop state.
Result
Visitors can understand and verify Dessi's work through more than one route, while Dessi retains control over which evidence becomes part of the public representation.
Boundary
DG-OS is functional as a public profile product. Automated private-to-public ingestion, durable activation, rollback, and the later-outcome feedback loop remain under development in Human Systems Platform.
DG-OS repository ↗
02 / CLAIMS

What the evidence supports.

Built and maintains the open-source AI News Hub as a shared research platform for applied AI in marketing, connecting news, deep analysis, build sessions, and public prototypes. Publicly verifiable

Visibility collaborative public

Verified 2026-08-23

Boundary The public site and repository verify the platform. Article bylines may represent collaborative newsroom work and do not establish sole authorship of every publication.

Created the platform to support an internal community of thinkers, builders, and innovators that has grown to about 400 company members. Owner-reported

Visibility collaborative public

Verified 2026-08-23

Boundary The community purpose and approximate membership are owner-reported. Public membership or engagement analytics were not available for independent verification.

Built AI Skills Platform to turn applied AI research into reusable capabilities that practitioners can inspect, install, test, and adapt across agent runtimes. Publicly verifiable

Visibility public

Verified 2026-08-23

Boundary The public catalog verifies package structure, tooling, and declared usability. Adoption and production impact require separate evidence.

I am developing Human Systems Platform as a founder-led product for people and organisations that need credible evidence of capability in AI-mediated work. It turns selected experience into private learning, owner-approved public evidence, and testable hypotheses about where that capability may create value. Owner-reported

Visibility public

Verified 2026-08-23

Boundary The shared product repository is private and the joined platform remains under development. Public component repositories support the direction, but product-market fit, beneficial outcomes, customer demand, and willingness to pay are not yet established.

I built Agentic Commerce to help brands test and improve how they describe products for paid placements and organic agent-led discovery, without giving an AI system unchecked control over business actions. Publicly verifiable

Visibility collaborative public

Verified 2026-08-23

Boundary The repository is an active public engineering project. It does not claim guaranteed production ranking outcomes or production-scale Codex traffic.

I designed a Bayesian-style learning loop that updates brand- and product-scoped beliefs as evidence arrives and carries supported patterns into later query and copy generation. It recommends whether to promote, revise, or reject each variant, while the supervised runtime keeps tools, evidence, memory, approvals, and recovery under explicit control. Publicly verifiable

Visibility collaborative public

Verified 2026-08-23

I built Learning Foundry as a stand-alone learning product for people who want to learn with AI without treating an agent's successful output as proof of their own understanding. It helps them explain, test, apply, and revise what they learn while separately developing evaluated agent capabilities. Publicly verifiable

Visibility submitted public

Verified 2026-08-22

Boundary Learning Foundry is a submitted OpenAI Build Week project. Its repository remains unchanged while judging is active.

I built DG-OS as a stand-alone public profile and discovery product where visitors can inspect a person's systems, evidence, writing, and development or ask questions grounded in reviewed public sources. Publicly verifiable

Visibility public

Verified 2026-08-22

I developed Gateplane independently to explore how authenticated human or agent identity becomes bounded, tenant-aware authority for tools and external effects. Owner-reported

Visibility public

Verified 2026-08-28

Boundary The product overview is public and the source is private. Gateplane is an independent beta, not an employer deployment, and production adoption is not claimed. LLM orchestration, governed learning, and production proof of the optional sandbox path remain incomplete.

Implemented behavioural embeddings, unsupervised clustering, and grounded persona generation for intent analysis. Publicly verifiable

Visibility collaborative public

Verified 2026-08-22

I architect and build the Programmatic plugin and agent harness that turns ambiguous user requests into tenant-bound execution across skills, a typed CLI, backend services, analytical data, and advertising-platform APIs. Owner-reported

Visibility private employer

Verified 2026-08-28

Boundary This is an owner-reported description of current employer work. Source code, client data, logs, infrastructure details, and operational measurements remain private.

I build backend ingestion and serving workflows for heterogeneous advertising data across provider hierarchies and analytical grains, and evaluate storage architecture against query shape, aggregation, latency, reliability, and cost. Owner-reported

Visibility private employer

Verified 2026-08-28

Boundary This is an owner-reported responsibility-level claim. Employer implementation details and measurements remain private, and active architecture proposals are not represented as deployed outcomes.

I delivered backend services and cloud data workflows for an award-recognised ecommerce optimisation platform. Owner-reported

Visibility private employer

Verified 2026-08-28

Boundary The public case study verifies the platform and award context. My detailed implementation contribution remains employer-confidential.

03 / EVOLUTION

How the work changed.

This record shows what changed, why it changed, and which questions are still open.

  1. revision

    Production and public evidence became separate proof planes

    Employer work now demonstrates production agent, data, and multi-tenant constraints at a confidentiality-safe level, while independent systems provide inspectable evidence of the corresponding authority, evaluation, and recovery patterns.

    reviewed
  2. revision

    The founder product became the orienting system

    Human Systems Platform now provides the product frame for Dessi Space, Learning Foundry, DG-OS, and the planned Organization Foundry. Each remains useful on its own while contributing to one owner-controlled evidence and learning loop.

    reviewed
  3. revision

    Research and implementation platforms became first-class systems

    The AI News Hub now appears as the shared research product, while AI Skills Platform appears as its practical capability layer for turning selected ideas into reusable packages.

    reviewed
  4. revision

    Current projects became one system map

    Dessi Space, Learning Foundry, DG-OS, and the shared Human Systems Platform are now described as separate authority planes with explicit dependencies, maturity, and publication boundaries.

    reviewed
  5. observation

    Models propose; validated systems decide

    Across Gateplane and Agentic Commerce, improving intelligence remains inside deterministic controls for identity, permissions, evidence, state transitions, effects, and rollback.

    observed
  6. revision

    Claims became typed evidence

    DG-OS replaced inherited profile content and broad self-description with claims that carry visibility, confidence, provenance, and an explicit boundary.

    reviewed
  7. observation

    Reliability is a product surface

    Across Agentic Commerce, Learning Foundry, and DG-OS, the recurring work is making agent plans, evidence, failures, and human decisions inspectable.

    observed
  8. question

    Can reliability become measurable across real coding tasks?

    The next research step is a reproducible task suite that records solve rate, regression, latency, token use, failure class, and intervention effect.

    active
  9. experiment

    Private evidence, deliberate publication

    Dessi Space now observes approved repositories read-only and records private review events. Public DG-OS receives only separately approved projections.

    active
Open Evidence & Evolution in DG-OS →
04 / BOUNDARIES

What the evidence does not establish.

  1. No public coding-task solve-rate benchmark yet.
  2. No claim of model fine-tuning experience.
  3. No public production measurements for token cost, latency, or capacity.
  4. Employer systems are described at outcome level; private code, logs, clients, and infrastructure remain private.
  5. Human Systems Platform is a founder product in private active development. Its components are functional at different levels, but the complete joined workflow and commercial demand are not yet proven.
  6. Current learning updates evidence, beliefs, memory, policy, and versioned harness configuration; it does not retrain a foundation model or permit silent self-modification.

Profile artifacts

Read the concise profile. Inspect the systems when detail matters.

Download CV · PDF DOCX Markdown Explore DG-OS