Application for ConductThe posting

ConductProduct EngineerLondon · 5 days/week on-site

Product Engineer · Conduct

Your agents map ERP customisation that somebody once wrote by hand, under deadline. I was that somebody — and I have spent this year building the agents instead.

Pharmacon is a turnkey ERP for one of Bangladesh's largest pharmaceutical plant-solutions providers: a client's plant requirement becomes international offers and quotations, becomes status-tracked negotiation, becomes local buying, and lands months later as a ledger entry. I was the hands-on full-stack lead. The client still runs it — last support round 2024. I know what the sprawl your agents read looks like from the inside, because I am the reason some of it exists.

Book 30 minutes

Ask about this section

Named on this page

NewScriber
Pharmacon
Elobooks & Anygum
Wiregent
Global Jute Trading
fauzul.com
french-with-fauzul

Two things Conduct needs at once, and I have both. I have built the kind of enterprise system your agents are pointed at — a turnkey pharma ERP, procurement through offers and negotiation to ledger, still in production seven years on. And I build agentically now rather than by hand: NewScriber runs ReAct reason-act loops with grounding evals gating every script against its sources before synthesis, and publishes to Spotify daily in four languages on that basis. React, Vite and TanStack are my daily front end. Kotlin and Temporal are the two honest gaps, and I have named them rather than let you find them.

Ask about this section

2024last support round on an ERP built in 2018

8days, scaffold to working platform

4×faster time-to-market

4languages published daily, evals-gated

Your posting, line by line

Every line Conduct posted, answered with an artifact.

Conduct's posting says language familiarity matters less than the evidence of what someone shipped. So every line below cites an artifact, not a familiarity.

Strong fit

You build with AI natively, not as an afterthought — your default reach is for the agentic solution, and you've got the receipts

Most of what I ship is directed rather than typed. Repeating workflows become reusable agent skills with explicit triggers so they fire reliably; hooks and scoped tool permissions give deterministic control at the points where a prompt alone burns budget; harness and model get routed per task rather than defaulting to one. Claude Code 101 and Introduction to Agent Skills from Anthropic sit behind it. The receipt is : scaffold to working platform in eight days, 195 commits, source files 101 to 222 and test files 17 to 47 — the tests outpacing the code is the part that only happens when the loop is set up properly.

Strong fit

You've shipped agentic systems and can point to how you handled the parts that don't work yet: failure modes, evals, tool design, harness behaviour

is the one I would open first. A cron-driven multi-agent pipeline parses uncached articles with Firecrawl, then runs ReAct reason-act loops to filter, rank and script — and grounding evals gate every script against the scraped source bodies before it reaches synthesis. Without that gate it is a demo; with it, it publishes to Spotify and Apple Podcasts daily in four languages. On tool design, Blur Studio runs one engine behind three callers — a web app for hand-painting, a CLI for pipelines, and an MCP server so an agent invokes the identical tool a human would, with no agent-only escape hatch. And 's model layer is vendor-agnostic and deterministic with no key and no network, so every path degrades rather than fails.

Strong fit

Obsessed with how AI systems behave, fail and improve — reliability, governance and evals as first-class engineering problems

Grounding evals are release gates on , not a dashboard. 's ledger is hash-chained so a diligence claim can be traced to the record it came from, and its model layer falls back deterministically rather than erroring when a provider is absent. On fauzul.com a 2B model — Gemma 4 E2B on WebGPU and MediaPipe, streamed from R2 — runs inside the browser tab so no prompt leaves the device. Governance has been a delivery constraint rather than a talking point: HIPAA and GDPR work, multi-tenant org and instance RBAC with SSO on Auth0, audit logging.

Ask about this section

How the harness behaves

What the harness owes the customer.

Evals as release gates

Generated output is checked against its sources before it reaches anyone. publishes daily on that basis; without the gate it would be a demo with a good screenshot.

One engine, three callers

Blur Studio exposes the identical tool to a web app, a CLI and an MCP server. An agent gets no private escape hatch, which means its behaviour stays debuggable by a human running the same command.

Degrade, never fail

's model layer is vendor-agnostic and deterministic with no key and no network. Enterprise reliability starts with what the system does when the provider is absent, not when it answers.

Hooks and scoped tool permissions

Deterministic control exactly where a prompt alone would burn budget — and the mechanism that lets an agent have real freedom without losing the guardrails an enterprise customer needs.

Ask about this section

The toolkit

Weighted to what this seat actually touches.

Weighted to what this seat touches: the agent layer, the React/Vite/TanStack front end, and the enterprise systems underneath.

Agentic systems & harness design

ReAct reason-act loops
Grounding evals as release gates
Reusable agent skills with explicit triggers
Hooks & scoped tool permissions
MCP servers
Claude Code subagents
Vendor-agnostic model layer with deterministic no-key fallback
Harness & model routing per task
Token-budget engineering
Anthropic SDK
Gemini
On-device inference (WebGPU)

React, Vite & TanStack

React 19
TanStack Start / Router / Query
Vite 6
TypeScript (strict)
Angular 14+
Zustand
Redux
Tailwind v4
Storybook design systems
Figma end-to-end
Playwright
Vitest
WCAG 2

Backend & long-running work

Node.js
Express 5
NestJS
Prisma
PostgreSQL (indexed, load-tested)
GraphQL Yoga with response caching
Go
Python
gRPC / Connect-RPC
Redis
n8n cron pipelines with retries
EventBridge / SQS / SNS
Zod contracts
Temporal — named ramp
Kotlin — named ramp

Enterprise systems

Turnkey multi-module ERP (Pharmacon)
SMB bookkeeping ERP (Elobooks)
Quotation & offer workflow
Status-tracked negotiation
Ledger & dues
Multi-tenant org & instance RBAC
SSO (Auth0)
Audit logging
HIPAA
GDPR
Reporting

Infrastructure & delivery

AWS Lambda incl. container images (early adopter)
ECS
S3 & MediaConvert
IAM & WAF
CloudFormation with SAML provisioning
Docker
GitHub Actions CI/CD (4× faster time-to-market)
Cloudflare Workers / R2
Cost-aware multi-region placement

Product engineering

Features owned end-to-end, inception through iteration
Speaking to enterprise customers directly
Product calls without full information
Outcomes over output
0→1 across five verticals
CSPO
Linear and Jira daily

Ask about this section

Built, shipped, still running

The builds behind every claim above.

Four builds that answer this posting directly — the agentic pipeline with its failure modes handled, the enterprise ERP written by hand, and the front end you already run.

01Builder · solo

NewScriber

A cron-driven multi-agent pipeline: Firecrawl parses uncached articles, ReAct reason-act loops filter, rank and script alternating dialogue, and grounding evals gate every script against the scraped source bodies before synthesis. Go and n8n over PostgreSQL, multi-speaker TTS, delivered to Spotify and Apple Podcasts on a regional schedule.

Go
n8n
PostgreSQL
ReAct + evals
Firecrawl
Docker

Why it matters hereThe answer to how you handle the parts that do not work — the eval gate is the only reason it can publish unattended, daily.

02Full-stack lead, hands-on

Pharmacon

A turnkey ERP for one of Bangladesh's largest pharmaceutical plant-solutions providers. A single deal starts as a client's plant requirement, runs through international principals for offers and quotations, continues into local buying, and lands months later as a ledger entry. Multi-module, RBAC, reporting, notifications.

Multi-module ERP
Quotation & offer workflow
Ledger & dues
RBAC
Reporting

Why it matters hereI authored the customisation sprawl your agents are asked to read — and supported it for years while the business could not stop running.

03Lead architect · Figma designer · shipped to production

Elobooks & Anygum

An SMB bookkeeping ERP running on a multi-tenant backbone I architected and designed myself: org-scoped and instance-scoped RBAC, per-instance isolation so one tenant's chart of accounts, ledger and permissions never reach another, and SSO on Auth0 across the whole estate. Built as an Nx monorepo on NestJS, PostgreSQL and Redis, with the API marketplace it sits inside shipped alongside it.

NestJS
Multi-tenant RBAC
Instance isolation
Auth0 SSO
PostgreSQL
Redis
Nx monorepo

Why it matters hereEnterprise ERP is a tenancy problem before it is an AI problem — instance isolation, org-scoped permissions and an audit trail are the primitives your agents have to respect when they change a customer's system.

04Founder · sole engineer

Wiregent

Diligence infrastructure for early-stage capital, built solo end to end on React 19, TanStack, Express, Prisma and PostgreSQL, with a hash-chained ledger and a vendor-agnostic model layer that is deterministic with no key and no network — every path degrading rather than failing.

TypeScript
React 19
TanStack
Vite
Express
PostgreSQL

Why it matters hereThe agentic receipt and the front end in one: eight days scaffold to platform, 195 commits, test files 17 to 47 while source went 101 to 222.

Ask about this section

Beyond the posting

What I would bring that Conduct did not ask for.

leverage

A CMS deleted from the architecture

A non-technical export team now ships content themselves: plain English in, reviewed pull request out, CI/CD deploys it. The win was removing a dependency the business had budgeted for indefinitely.

Global Jute Trading

reliability

A 2B model inside the browser tab

Gemma 4 E2B on WebGPU and MediaPipe, streamed from R2 behind a custom token-streaming renderer. No prompt leaves the device.

fauzul.com

enterprise

An ERP still running seven years on

Procurement through international offers, negotiation states and local buying into the ledger — built 2018 to 2019, last support round 2024, still the system the business runs on.

Pharmacon

consistency

977 consecutive days of French

Published grammar explainers, not a streak screenshot. A metric nobody asked for, which is rather the point.

french-with-fauzul

Ask about this section

The call, started early

The questions a Conduct engineer would actually put to me.

You've never written Kotlin or used Temporal. Why is that not disqualifying?

Because your own posting says language familiarity matters less than the quality of the thinking and the evidence of what was shipped, and I would rather be measured on that. Both gaps are narrow and I have named them upfront. Kotlin sits on habits I already have — strict typing, null-safety, structured concurrency — with the JVM ecosystem as the genuine curve. Temporal formalises something I have hand-rolled: cron-driven multi-agent workflows with retries and partial-failure handling on , EventBridge and SQS/SNS in production. What is not a ramp is the agentic practice, the React/Vite/TanStack front end, or having built the kind of ERP your agents are pointed at.

Everyone says they build with AI now. What makes your version different?

That I build the scaffolding rather than the prompts, and that I can show you where it fails. Repeating workflows become reusable skills with explicit triggers so they fire reliably; hooks and scoped tool permissions give deterministic control exactly where a prompt alone would burn budget; harness and model route per task. Then the reliability layer: grounding evals as release gates on , a vendor-agnostic model layer on that is deterministic with no key and no network, a hash-chained ledger so a generated claim traces to its record. The measurable version is — eight days scaffold to working platform, 195 commits, test files 17 to 47 while source went 101 to 222.

You've been a CIO and a founder. Are you going to be happy as one of thirty, owning a domain?

I stepped out of the operating role in August 2026 specifically to get back to building, and everything since has been hands-on — sole engineer on , sole engineer on . I was an IC first as well, through Principal Staff Software Engineer, before I managed anyone. Thirty people in one London room shipping into DHL and Porsche is the density I left the CIO seat to get back to. Owning a domain end to end, with the customer attached, is the job I want rather than a step down from one I had.

You've built ERPs, but not SAP. How much of that actually transfers?

The problem shape transfers; the vocabulary does not, and I would not blur those. What transfers is having authored the sprawl: on Pharmacon I wrote the custom modules, the negotiation states, the ledger logic and the dependencies nobody wrote down, then supported it for years while the business could not stop running. That is the exact artifact your agents are asked to read, and knowing why it looks the way it does is useful when you are deciding what an agent should infer versus ask. What does not transfer is ABAP, S/4HANA clean-core rules and SAP's own idiom for customisation — that is a real curve, and the fastest way through it is customer calls rather than documentation.

How I would start

Give me a customer domain in week one with the whole problem attached — the enterprise on the phone, the harness underneath, and the accountability for whether it works.

The working model

The reservation

The one Dex put in writing: five days a week in the office, against a recent background of remote founder and advisor work.

My answer

It is the right question and the answer is straightforwardly yes. The remote years were a consequence of running a distributed firm, not a preference I fought for — and the two things I have missed most since stepping out of the operating role in August 2026 are a room and a whiteboard. Conduct is thirty people in London shipping into DHL and Porsche from the same building; being in that room is most of why the role is interesting. London is also the easiest city on my list to land in: my sister lives there, it has been a delivery geography of mine for years, and UK Global Talent is self-sponsored and already progressing independently of any single role. I am not asking to negotiate the five days down.

London is the easiest city on my list to land in, and the pull is genuine rather than logistical: one of Europe's strongest technical hubs, a delivery geography of mine for years, and my sister lives in London, so it comes with people rather than just an address. UK Global Talent is self-sponsored and already progressing independently of any single role; Conduct's sponsorship or a Skilled Worker route work equally well. Five days on-site is a draw rather than a concession — the room is the point.

Happy to walk Conduct through any line of this.

Book 30 minutes

Ask about this section