/// Hire

Open to roles

One page of resume.
A site full of proof.

Applied AI engineer · forward deployed · client delivery. I build agent systems and the machinery that proves they work — deploy gates, evidence ledgers, rollback paths. Status labels below are literal — shipped means shipped, and a prototype says so.

Drop me into a team whose agents work in the demo and break in production, and in week one the failure is surfaced, gated, and rolled back — not argued about.

abdur.ai/hire · US Letter · one page · ATS PDFs regenerated 2026-10-06

Applied AI & Systems Engineer · Agent Orchestration & Verifiable Execution

Abdur Rahman M. Sayeed

Prospect Park, NJ · (201) 321-2235 · asayeed95@outlook.com
linkedin.com/in/asayeed95 · github.com/asayeed95 · portfolio & hiring profile: abdur.ai/hire
U.S. work authorized · no sponsorship required · remote / NJ·NYC / relocation

abdur.ai/hire

Applied AI engineer and client-delivery systems builder. I enter real businesses, learn the operator constraints, and build the systems underneath them — agent orchestration, prompt and system design, browser automation, and AI-assisted development, on a practical enterprise-IT foundation.

/// Experience

Applied AI & Systems Engineer

2020 – Present

One Asec LLC · independent practice since 2020, incorporated 2021 · Prospect Park, NJ / Remote

Client engagements below are delivery phases, most recent first.

Applied AI Products

Sep 2025 – Present
  • Architect systems, orchestrate AI coding agents to implement them, and audit every change for verifiable execution.
  • Northsun — AI-agent infrastructure with governed context and evidence systems. HeyCLI — voice and coding-agent orchestration across sessions.
  • BrowseFlow — accessibility-tree browser automation with approval points. Contributor on Dockerfile.ai.

Enterprise Retail IT Delivery

Apr 2024 – Jul 2025
  • Independent IT contractor for ETRO Fashion US (client) via Vox Elements, billed through One Asec — 200+ users, 14 sites, 500+ tracked assets.
  • Identity and endpoint policy across Active Directory, Entra ID, and Intune; Lansweeper, Cisco/RingCentral, Jira triage, cross-site escalation.

Automotive Business Systems & Growth

2020 – 2024
  • Built business, IT and financial infrastructure from the ground up for Autoblitz LLC (NJ), Carfixed LLC (NJ) and Exquizit Auto Repair (Mount Vernon, NY): entity setup, networks, diagnostic workstations, POS and inventory.
  • QuickBooks, Excel, and Power BI reporting for pricing, profitability, and cash flow; reconciliation automated with Python and SQL.

/// Earlier experience

Financial Operations & IT Specialist

Jan 2019 – Jan 2020

Vital Life · Ottawa, ON

Owned bookkeeping, budgeting, KPI, CRM, invoicing, and monthly reporting while expanding the company's marketing and IT foundation.

/// Education & Certification

University of Ottawa — B.S., Financial Mathematics & Economics (2015–2019) · Coding Dojo / Colorado Technical University — Cybersecurity Certificate, 24-week intensive (2023–2024) · CompTIA Security+ (2024)

/// Selected systems

Northsun

Northsun is the memory and enrichment layer for AI agents. Gate-first governance; bi-temporal evidence on every fact.

HeyCLI

Voice multi-session orchestrator routing work to coding agents via a Node/TS bridge and iOS client.

BrowseFlow

Agent-agnostic browser automation — accessibility-tree perception, approval queues, run evidence.

Dockerfile.ai

Dockerfile generation, explanation, and optimization — verified in a sandbox before you trust it.

/// Core skills

Applied AI — AI system design, prompt engineering, agent orchestration, workflow & tool design, browser automation, AI-assisted development

Client delivery — customer discovery, requirements translation, technical delivery, product judgment, operator empathy

Engineering — TypeScript, React/Next.js, Node.js, Python, SQL, Git, API integration, UX/product design

Operations — Microsoft 365, Active Directory, Entra ID, Intune, Jira SM, Power BI

/// What I build

Systems that show their work.

I work the forward-deployed pattern: sit with the operator, learn the real constraint, build the system underneath it, then build the evidence that proves it behaves. Most of my recent work is agent infrastructure — memory and governance for AI agents, browser automation with approval points, and multi-session orchestration for coding agents.

evidence over assertion

Decisions land as dated records with the option chosen, the alternatives rejected, and the rollback path. A locked record is never edited in place — amendments get their own dated file, so the history stays readable.

gates that fail loudly

Continuous deployment queries the production schema before it will deploy, so a migration that did not actually land stops the release instead of silently passing. Every migration ships with a down-migration and a runbook.

agents that review agents

I run multiple frontier models against each other with adversarial review before merge, and treat their output as untrusted until a gate confirms it.

/// Corrections

Twelve reviews to learn one thing.

I wrote an operating charter for a fleet of agents and put it through twelve independent adversarial reviews — each a fresh context, each on a different provider from the one that wrote the document. Substantive findings per pass:

17
17
19
15
22
9
14
4
7
7
10
12

substantive findings · pass 1 → 12

It never converged, and the reason was the finding. Every surviving objection had the same shape. A daily cap that prose “enforces” is not atomic. A safety check that must run “immediately before sending” cannot run inside a third-party scheduler. “The author is never the approver” cannot be proven while every lane commits under one identity.

Prose cannot enforce a mechanism. So I stopped rewriting the document and took the authority out of it: every outward action became frozen by default, and the permission table moved into code, where a uniqueness constraint can hold what a sentence cannot. The document got shorter and the system got safer.

The audit that made the point concrete: I inventoried the automations across five roots and found that none of them had an independent witness. Every one could have died silently and reported nothing — an unwatched automation is indistinguishable from a dead one. Witness-before-workload is now the first thing I build on any pipeline, ahead of the pipeline.

/// How I work

Many agents, one accountable lane.

I run coding and research agents from four providers concurrently — Claude, GPT, GLM, Kimi — across several machines, with cross-session message passing and a shared decision log. It is not a novelty setup; it is how the work gets done. What makes it survivable is the operating model underneath it.

one mutable owner

Two agents editing one artifact is a merge conflict with extra steps. Every work item names its owning lane before a byte is written, and a second agent reviews rather than edits.

the author never approves

Every change is reviewed by an agent on a different provider with fresh context. Same-model self-review finds what the model was already inclined to miss. Law-touching changes always require a non-author reviewer; there is no self-review exemption for them.

verify the substrate, not the name

A review is independent only if the runtime actually differs. I check that, because a "second reviewer" running the same model underneath is a rubber stamp with a different label.

/// Client delivery

Inside someone else's constraints.

Enterprise retail IT delivery

Apr 2024 – Jul 2025

At One Asec, delivered a retail IT subcontract for Vox Elements supporting ETRO Fashion US — 200+ users, 14 sites, 500+ tracked assets. Identity and endpoint policy across Active Directory, Entra ID, Intune; Lansweeper, Cisco/RingCentral, Jira triage, cross-site escalation.

Automotive business systems

2020 – 2024

Built and operated business, reporting, and IT foundations for three automotive-service businesses. QuickBooks, Excel, Power BI; reconciliation automated with Python and SQL.

/// Selected systems

Five systems, labeled honestly.

Status labels are literal. Shipped means shipped; a prototype says so.

Northsun

In development · SDK on npm · not GA

Northsun is the memory and enrichment layer for AI agents.

Relay

Deployed pilot · agent control plane in review

Event spine for the company's automations: authenticated webhook ingress, fenced leases, receipts that make a duplicate impossible.

BrowseFlow

Working prototype · reproducible locally

Agent-agnostic browser automation — accessibility-tree perception, CDP-first hybrid engine, approval queues, run journal.

HeyCLI

Working prototype · iOS client + bridge

Voice-driven multi-session orchestration for coding agents over an authenticated WebSocket.

Dockerfile.ai

Shipped

Dockerfile generation, explanation, and optimization — verified in a sandbox before you trust it.

/// Northsun — architecture

Memory as governance, not a bolt-on.

Northsun is the memory and enrichment layer for AI agents. Mnemix is the Memory Lab / Forgetting Test — a free diagnostic from Northsun.

Gate-first, router-second. The architecture is designed so no write lands until it clears the gate, and corrections deprecate old beliefs rather than silently overwriting them. BEAD is the spine: two clocks on every fact — when it was true, and when it was learned — so a retrieval can be replayed and graded against reality.

voice · telephony

inbound callers, live turns

agent runtimes

tool calls, session events

app events

webhooks, product signals

▼  observe — throttled, durable, replayable intake

the gate

gate-first · router-second

Designed so no write lands until it clears the gate. Corrections deprecate old beliefs — nothing is silently overwritten.

▼

BEAD — bi-temporal, evidence-anchored decisions

the memory spine

Two clocks on every fact — when it was true, when it was learned — so any retrieval can be replayed, audited, and graded against reality.

Cloudflare WorkersHonoSupabase · Postgres · pgvectorUpstash Redis + QStashSentryMCP SDK

read path — returns enriched

recall

POST /v1/recall_and_enrich

The right memory for this moment, already enriched and ranked.

enrich

Trestle · Twilio Lookup

A caller resolved into real context before the agent ever speaks.

Only shipped endpoints are listed. Anything still in development stays off this page until it deploys.

Counted in the repository · main · 2026-09-03

330

test files · *.test.ts across the workspace

15

CI workflows · every merge runs them

16

workspace packages

@mnemix-ai/client

published SDK on npm · v0.2.2

/// Case study — content pipeline

Design complete · four independent reviews · build in progress

The hard part of a content pipeline is not the writing.

An event-driven pipeline that turns merged engineering work into published writing: extract the evidence from a merge, draft against a claims policy, review on a different provider, publish, measure, and feed the result back to whoever writes next.

Generation is the easy half. The hard half is that publishing is an outward action, so every stage needs a lease, a receipt, an idempotency key, a kill switch, and a named human gate — and none of that can live in the prompt.

The specification went through four independent adversarial reviews before a line of it was built. The third one found that a piece could be corrected after approval while its already-scheduled derivatives kept the old version queued to fire. That is a defect you find on paper, or you find in public.