ARUM MOHEE · AI PRODUCT ENGINEER · MANCHESTER

I build AI systems people can actually use.

I work inside agencies and enterprise teams to turn operational problems into governed AI products. I map the workflow, engineer the agent and evaluation layers, design the interface and embed the result into the way the team already works.

See how I build ↓Discuss the role

EnterpriseWorkflow discovery through adoptionEngineeringAgents, evaluation and integrationProductInterfaces people can learn and trust

LIVE OPERATION / CAMPAIGN DELIVERY RUNNING

NOWresearch_agent / sources verifiedhuman owner: confirmed

COMMERCIAL CONTEXT / BUILT THROUGH DELIVERY

Nearly $0 million

Client revenue supported

Across campaigns, systems and operating work I have contributed to.

0+ brands

Worked across

Ecommerce, agency delivery, creative production and product work.

0 disciplines

One accountable operator

Strategy, engineering, product interface and adoption under one owner.

ENTERPRISE AI DELIVERY / EMBEDDED PRODUCT ENGINEERING

From operational problem to production AI capability.

I can enter an ambiguous workflow, find where intelligence creates real leverage and carry the work through architecture, engineering, interface design and adoption. The model is one component. The governed operating system around it is the product.

Discover the real constraint

Map decisions, hand-offs, data, permissions, failure points and the work people quietly perform outside the official process.

WORKFLOW · RISK · VALUE

Architect the intelligence layer

Define specialist agents, context boundaries, memory, model routing, tool contracts and the state that must persist between steps.

AGENTS · RAG · MCP · STATE

Engineer trust into the workflow

Use structured outputs, evaluation suites, source gates, scoped credentials, approval lanes and reversible actions before anything reaches production.

EVALS · POLICY · HUMAN GATES

Embed, measure and improve

Ship inside the tools the team already uses, make progress observable, train the operators and retain corrections as controlled memory.

INTEGRATE · ADOPT · LEARN

PRODUCTION CONTROL PLANE / LIVE GOVERNED

01 / CONTEXTBusiness truthbrief · policy · historyAccess boundaryidentity · permissionsSource evidenceretrieval · citations

02 / AGENT RUNTIMEResearchfind · compare · verifyReasoningplan · route · decideProductionbuild · transform · act

03 / VERIFIER GATESSchema + rulesdeterministic checksEvaluationquality · risk · latencyHuman authorityapprove · revise · stop

04 / ADOPTIONTeam interfaceclear state + controlExisting stackAPI · Slack · workspaceOperating memoryoutcomes · corrections

01Business context firstTechnology follows the workflow. 02Agent proposesFlexible reasoning prepares the work. 03Verifier gatesRules, tests and humans decide what proceeds. 04Integration is the productCapability lives where work already happens.

WHY I AM BUILT FOR THIS ROLE

I don’t wait for the AI landscape to settle.

I maintain the systems that tell me what changed, test what matters and turn useful advances into working capability before most teams have finished discussing them.

LIVE AI TEST LAB / 07:42 CONTINUOUS R&D

INCOMING RELEASESeedance 2.0Multi-shot video continuity update

CHANGELOG READResearch agent

REAL WORKFLOWRun in isolation

QUEUED

EVALUATIONQuality, control and latency

WAITING

HUMAN REVIEWUseful enough to adopt?

WAITING

OUTPUT QUALITY—

CONTROL—

LATENCY—

TESTING

MODEL WATCH

Gemini Live APIClaude tool useCodex workflowHiggsfield Cinema StudioVeo evaluationKling update

SELECTED CASE / AGENCY AI OPERATING SYSTEM

One system. Every hand-off visible.

My strongest work is not a collection of prompts. It is a connected operating model for research, creative decisions, production, approval and learning—with specialist agents supporting each stage and humans retaining authority.

LIVE CASE / SPRING CAMPAIGNFrom brief to accountable delivery

01 · INTAKE

BRIEF RECEIVED

The system begins with the real commercial constraint.

Product facts, audience, budget, deadline, owner and approval route enter together—before an agent generates anything.

OwnerMarketing leadDeadline14 days

CAMPAIGN BRIEF / V4OWNER CONFIRMED

Launch the spring range without losing brand trust.

Objective: Qualified demand
Audience: Existing + new
Deadline: 14 days

INTAKE COMPLETE

VOICE OF CUSTOMER184 signals clusteredReviews · calls · search

COMPETITOR FIELD12 patterns mappedClaims · offers · creative

PRIOR PERFORMANCE6 learnings retainedEvidence linked

24 verified sources

APPROVED
TRUTH

ROUTE AAuthority

ROUTE BMomentum

ROUTE CProof

Three genuinely different routes. One controlled brief.

DECISION ROOM / ROUTE B

Direction ready for accountable approval.

Evidence7 sourcesBrand riskLowOwnerMarketing lead

HUMAN
DECISION

delivery.run

01 Construct campaign assets DONE

02 Validate claims and links DONE

03 Render responsive variants DONE

04 Route to accountable owner RUNNING

18 / 18
checks

RESULT
+18%

RETAINWinning angle

CORRECTRejected claim

IMPROVENext brief

OutcomeCorrectionControlled memory

Context attached

SELECTED PRODUCT WORK

Proof across realtime, native and agentic software.

Four products are enough to show the range: a live collaboration environment, local voice software, an escalation system and encrypted iMessage media.

Draw directly on the workPen, arrow, rectangle, highlighter, named cursors and comment pins.

Replay the exact momentReview the session, export local clips and retain full-resolution snapshots.

Captions and transcriptSpeaker-aware captions, transcript download, activity and AI meeting notes.

Mara follows the actual callPrivate voice or text answers grounded in recent conversation and screen context.

Team and Sales modesMove the work together or surface talk share and live coaching signals.

01 / 07One room for the live workVideo, shared context and the working surface stay together.

MARKER / AI-NATIVE REALTIME PRODUCT

The meeting becomes a shared working surface.

I designed Marker around the work people normally lose between a call and the follow-up: live context, visual feedback, participation signals, decisions and the action that follows.

Collaborate liveShare, point, draw, box and comment directly over the work.
Replay and retainReview the session, preserve local clips and full-resolution snapshots, then leave with the thread intact.
Understand the roomSpeaker-aware captions, transcript, talk balance, decisions and AI meeting notes.
Act with contextMara answers privately against the current conversation and structures follow-up.

LiveKitDeepgram Nova-3Gemini LiveAnthropicNext.jsElectron

SHIPPED macOS UTILITY

Gobby

Hold Option, speak and release. Audio and transcription stay on-device; a resident Whisper server keeps warm inference near-instant.

~0.05–0.07s: warm inference
Local: audio + transcription
Universal: Apple Silicon + Intel

Audio never leaves the Mac
Warm model avoids reload delay
Types into any active application

notes.txt

LISTENING ON DEVICE

Speaking…

WORKING iOS PROTOTYPE

Sentinel

Important work should not disappear inside ordinary notifications. Sentinel observes selected sources, classifies urgency and follows a configurable escalation policy.

GmailCalendarSlack→Urgency + policy engine→ReminderSMSCall

UNDERSTANDWhat needs attention?

Extract the commitment, deadline, sender and consequence—not just the notification text.

DECIDEWhat should happen next?

Apply owner rules, quiet hours, confidence thresholds and a timed escalation ladder.

PROVEWhy did it escalate?

Keep the source, reason and action visible so every automated step can be reviewed or reversed.

Sentinel iOS prototype showing time-based escalation tasks

NEXT ACTIONClient approval unansweredCall owner in 7 minutesreason attached · reversible

FIELD-TESTED / TWO IPHONES

SnapMessage

A view-once photo experience inside iMessage: capture, encrypt on-device, relay through expiring storage and burn after viewing.

SwiftUICryptoKitVercel Blob

ECOMMERCE CREATIVE SYSTEM

Listing Studio

Agents turn product facts, brand language and references into a controlled production brief, then route each asset through isolation, reference locking, campaign generation and visual QA.

Research brand + product truth
Lock claims, pack geometry + references
Generate listing masters + campaign worlds
Compare, verify + approve before export

Clean white-background marketplace product master — 01 / MARKETPLACE MASTER

AGENT-SUPPORTED GENERATIVE PRODUCTION

Brand intelligence in. Directed film out.

Research agents gather the brand, product facts, audience language and visual codes. A model-selection workflow routes each shot to the right engine; reference systems protect continuity; human direction controls story, camera, edit, sound and final approval.

Higgsfield

Veo 3.1

Imagen 4

Gemini Flash Image

Nano Banana Pro

Kling 3.0

Seedance 2.0

GPT Image

Sora

DaVinci Resolve

TECHNICAL RANGE

Enough depth to design the system. Enough product judgement to make it usable.

Agentic engineering

Design persistent specialist agents with scoped roles, context, memory, tool access, evaluation gates and accountable human escalation.

Claude Code · OpenAI Codex · RAG · MCP · context engineering · orchestration · structured outputs

AI product engineering

Turn capability into interfaces and services people can understand, operate and trust from prototype through deployment.

TypeScript · Node.js · Python · React · Next.js · Swift · SwiftUI · Electron · APIs · Vercel

Realtime + voice

Build low-latency rooms, streaming transcription, conversational assistance and private on-device voice workflows.

LiveKit · Deepgram · Gemini Live · Anthropic · whisper.cpp · local inference · event streams

Evaluation + governance

Define the evidence, tests, permissions and human decisions that make agent behaviour inspectable and safe to operate.

Evaluation suites · source gates · policy checks · observability · dry runs · audit logs · rollback paths

Generative production

Engineer model selection, reference locking, shot continuity, visual QA and finishing pipelines for controlled creative output.

Higgsfield · Seedance · Kling · Veo · GPT Image · Gemini · DaVinci Resolve

Business integration

Embed capability inside the tools, approvals and operating rhythms the business already depends on.

Slack · Notion · Figma · Klaviyo · Shopify · APIs · webhooks · scheduled jobs · operating memory

AI ENGINEERING / PRODUCT / TRANSFORMATION

Bring me the workflow everyone has learned to tolerate.

I’ll show you what should remain human, what AI can carry and how to make the result usable.

01Enterprise transformation02Embedded agency capability03AI product engineering