Systems Architect · Enterprise & Agentic AI
Arslan Khan
A systems architect for the Fortune 500. For sixteen years I’ve modernized legacy platforms and turned complex, business-critical systems into ones that are simple, dependable, and built to scale. Today that work is AI transformation, enterprise AI engineering, and agentic systems built for production, not demos.
System Architecture
Event-driven, distributed systems built to scale.
Legacy Modernization
Aging platforms transformed without the big-bang risk.
Enterprise AI
Production AI engineered into regulated, real operations.
Agentic AI
Autonomous agents that take action, not just answer.
Trusted with software for
National platforms for Brunei Darussalam.
Government-grade engagements I led, under official-secrecy obligations. Shown at a high level, the engagement and the responsible authority only. Detail under NDA.
Multi-agency governance · sovereign cloud
Wider record across ASEAN, US & South Asia
Sixteen years of enterprise systems, architected to last.
Enterprise clients
- Head of EngineeringNCR Voyix
Retail · Hospitality · Finance
Led 12+ engineers modernizing fragmented legacy into one platform, onboarding that cut client setup time 50%.
- Principal ArchitectSCTE · ISBE
Telecommunications
Re-architected a 15-year-old system into event-driven serverless, infrastructure cost down 70×.
- Chief Technology OfficerWeb Shop Manager
Automotive e-commerce
Full-stack revamp across a 10M+ product catalogue, real-time analytics and a 95/100 performance score.
- Principal EngineerThe New York Sun
Digital media · publishing
A complete digital transformation of the publishing platform, rebuilt end to end.
- Principal AI ArchitectLufthansa Group
Aviation
AI-native, multilingual, accessibility-first passenger voice platform, live at Group scale.
How I work
Simplify the complex
Strip a problem to its essence, then make the one strategic technology decision that drives a concrete result.
Transform, don't rewrite
Modernise legacy at scale with the strangler pattern, continuity first, no big-bang risk.
Event-driven & serverless
Distributed systems on messaging queues and serverless compute, 15M+ operations at minimal cost and downtime.
Ship with discipline
CI/CD, automated testing, and continuous monitoring, reliability engineered in, not bolted on.
07 engagements · 06 industries
Select AI Systems
Seven AI engagements I designed, built, and shipped across six industries. Each one a working system in production, not a demo.
Click an engagement to jump in · arrows walk the full sequence
Mediyn
Private clinical AI that turns a therapy session into documentation, insight, and homework, with PHI redacted on-device, before anything leaves the room.
PHI stripped on your device. Before anything uploads.
THERAPISTLast week we talked about the review at work. How has it been sitting with you?
CLIENTRough. I barely slept on TuesdayDATE. The meeting with DanielNAME kept replaying in my head.
CLIENTThat I'd freeze again, like at the HartwellORG office back in MarchDATE. My chest gets tight just thinking about it.
CLIENTI almost called in sick. MayaNAME told me to push through, but I couldn't focus all day.
Rough. I barely slept on [DATE]. The meeting with [NAME] kept replaying in my head.
That I'd freeze again, like at the [ORG] office back in [DATE]. My chest gets tight…
- ✓Raw audio never uploaded to any server
- ✓The cloud only ever sees de-identified data
- ✓All 18 HIPAA identifier categories stripped
Mediyn writes the note, picks the assessments, and builds the homework.
Subjective
Anticipatory anxiety and disrupted sleep around workplace evaluation.
Objective
Anxious but engaged; no suicidal ideation reported; safety screen unremarkable.
Assessment
Elevated generalized anxiety; avoidance urges present.
Plan
Weekly CBT. Thought-record homework; administer GAD-7.
GAD-7
86%Generalized anxiety
PHQ-9
61%Depression screen
PCL-5
34%PTSD checklist
PDFThought Record
Week 3 · prepared from this session
Situation
Presenting in the weekly review
Automatic thought
“I’ll freeze again.”
Evidence for / against
Prompted from session content…
→ Delivered to the patient portal
It knows the patient, not just the session.
Longitudinal clinical intelligenceSixteen to seven on the GAD-7 in eight weeks, and the system saw it happen.
Risk screened from day one
C-SSRS auto-assigned at intake: safety flags surface before they become emergencies.
Insights cross-referenced
Session 4's “avoidance urges” insight is read against the trend: interpretation, not just numbers.
Crossed into Mild in week 5
The severity shift is detected automatically and carried into the next note and recommendation.
PHQ-9 and GAD-7 trends update automatically·Session insights inform assessment interpretation
Every session makes the next one smarter.
Homework feeds back in.
Completed worksheets inform the next session's note and the next recommendation. Nothing starts cold.
Trends update themselves.
PHQ-9 and GAD-7 refresh automatically. Assessment data is cross-referenced with every session's insights.
Context carries across care.
Notes, instruments, and worksheets share one longitudinal picture of the patient.
Private clinical AI. Designed, built, shipped.
Everything on these screens is engineered by Arslan Khan: the on-device redaction engine, the clinical synthesis pipeline, the assessment intelligence, and the compliance envelope around them. Running in production. Session ends, notes done, worksheet delivered.
2-3min
to review and approve, per session
up to10hrs
back in the clinician's week, documentation and admin
21
validated instruments, on every plan
HIPAA · BAA on every plan · SOC 2 Type II · AES-256 · 7-year immutable audit
Watson
An autonomous AI voice agent that answers, books, and triages every inbound call, 24/7. Here, Watson is the product.
The production agent’s own voice and script, as deployed
7:42 PM. The office is closed. Watson isn’t.
Answers · books · triages · 24/7RIVERBEND THERAPY
MAIN LINE · 7:42 PM
Thanks for calling Riverbend Therapy, this is Watson. How can I help?
I need to move my Thursday appointment, something came up at work.
I see Thursday 4:00 PM with Dr. Reyes. I can offer Friday 11:00 AM or Monday 2:30 PM.
Friday works.
Done. Friday 11:00 AM. I've texted you the confirmation.
Live transcription · recording disclosed
Answered on the first ring · recording disclosed
Intent classified mid-call
Calendar lookup: Thu 4:00 PM · Dr. Reyes
Offered alternatives: Fri 11:00 · Mon 2:30
Caller confirms Friday
Calendar updated · confirmation SMS sent
SMS → caller’s phone
Riverbend Therapy: You’re confirmed for Fri 11:00 AM with Dr. Reyes. Reply C to cancel.
Real actions in real systems, mid-call
Everything your front desk does. Handled.
Live product panels: calendar, routing, answers, leads, analytics
Books & reschedules
Locks in the appointment before the call ends.
✓ Booked by Watson · confirmation sent
Triages & routes
Urgency assessed; every call sent to the right place.
Answers questions
Hours, insurance, pricing. All from the practice’s knowledge base.
Captures every new-patient lead
3 todayName, number, reason: a ready-to-work lead while you’re in session.
Maria Alvarez
Anxiety · trouble sleeping
David Cho
Couples therapy
Priya Nair
New patient · daytime availability
Logged, transcribed, measured
Every call summarized, with the numbers to prove it’s working.
182
this week · 0 to voicemail
✓ 100% transcribed
✓ 41 booked
A production voice agent, built end to end.
Telephony, real-time speech, intent classification, live calendar and EHR actions, guardrailed behavior. All engineered and operated by Arslan Khan. The economics work because the engineering does: one saved call pays for a year of it.
0
calls sent to voicemail with Watson
24/7
answered: nights, weekends, overflow
$2,500+
typical value of one new client*
Lifetime value varies by practice and payer mix
Talent Genius
AI embedded in the hiring pipeline of Banque Saudi Fransi: applicant screening, ranking, and first-level interviews conducted by an AI, on video.
The first interview is conducted by an AI, on video.
✦ AI first-level interview · asyncIn your IFRS 9 implementation, which methodology choice drew the most challenge from audit, and how did you resolve it?
“The SICR threshold for our SME book. Auditors pushed back on our 30-day backstop versus relative PD movement. We ran a 12-month look-back showing the relative-PD approach captured ~84% of eventual stage-3 migrations…”
✦ AIDemonstrates depth on a contested methodology point. Quantified support; recognizes the audit perspective.
Overall
0/100
81
CONFIDENCE
STRONG
89
COHERENCE
STRONG
76
SPECIFICITY
MODERATE
✦ Clear, structured answers; minimal filler.
✦ Strong IFRS 9 fluency; solid sector frameworks.
✦ Structured under pressure on the case prompt.
✦ Values stewardship; mentions team coaching twice.
✦ Wholesale credit + IFRS 9 directly relevant.
06:42: long pause (8s) on structured finance·19:11: hedged answer on RE sector outlook
✦ AI RECOMMENDATION · ADVANCE TO PANEL
88%Domain depth consistently above bar. Real-estate structuring is the stretch area for the panel to probe. Keep it in scope.
Talent Genius reads every CV like a credit officer.
✦ Scored against the requisitionKhalid Al-Mansour
Credit Analyst · ANB · 5 yrs wholesale credit
AI match vs requisition
0%
✦ ADVANCE TO INTERVIEW
92%Strong wholesale credit exposure with hands-on IFRS 9 ECL. Both skill gaps addressable via BSF Credit Academy.
Skills · experience · cultural fit · risk flags, all from the CV alone
✓Led IFRS 9 ECL rollout for ANB wholesale book
✓CFA Level II completed (May 2024)
✓Strong Arabic regulatory drafting
!Limited structured finance / syndications exposure
!60-day notice period vs target start
Q1Walk us through a credit decision where you overrode the model output. What signals tipped you?
Q2Describe your role in your bank's IFRS 9 ECL implementation. Which methodology choice was most contested?
→ HANDED TO FAISAL · Q2 IS ASKED AT 06:14 IN THE INTERVIEW
Before any human looks, Talent Genius has ranked every applicant.
Hessa Al-Fahad
Auditor · KPMG
Mohammed Al-Otaibi
Junior Analyst
Omar Hadid
Treasury Ops
+35 more · all screened by AI
Yousef Al-Sharif
0%
Credit Analyst II · GIB
TOP MATCHReem Al-Mutairi
0%
Wholesale Risk · Riyad Bank
HIGH POTENTIALAhmed Al-Rashid
0%
Senior Analyst · Alinma
STRONG MODELINGSara Al-Zahrani
0%
Credit Analyst · SNB
TAMHEER '24Noura Al-Dossari
0%
Risk Officer · BSF
INTERNAL MOBILITYKhalid Al-Mansour
0%
AI screening · scored 84/100
ADVANCE TO PANELEnterprise AI, delivered inside a major bank.
Talent Genius and Onboard Wiz: AI screening, ranking, and first-level video interviews embedded in Banque Saudi Fransi’s hiring pipeline. Engineered by Arslan Khan to institutional standards: bilingual, auditable, and structured for a regulated environment.
3
AI systems in one pipeline: screening, question generation, video interviews
5
evaluation dimensions, scored and logged on every interview
EN·عربي
bilingual by design, for a Saudi institution
Lufthansa
AI-powered, accessible, multilingual passenger experience, at production scale across the Lufthansa Group.
The whole journey, booked in one conversation.
AI-native voice platform · with Digital HangarHi Clark, I've booked Four Seasons for 3 nights while you're in Berlin.
Voice · natural dialogue, not rigid forms
✓Your ticket booked
MARKUS JENSEN
AU 199 · SYDNEY → TOKYO
New York
13 MAY - 12 JUNEPlaza City Gardens
CONFIRMEDProactive · before you ask
Hi Jane, your flight will be delayed by 3hr : 20min. I suggest the VIP lounge to lay back.
Same platform · Español
Hola John, tu vuelo se retrasará 3 horas y 20 minutos.
Built so every passenger can use it.
Accessibility was the primary gap: screen-based booking excluded passengers with disabilities. Voice closed it.
Universal accessibility.
Equal ease of use for blind and specially-abled passengers. Voice replaces the screen entirely.
Every language the Group flies.
Multilingual AI across Lufthansa Group's global passenger base.
Intelligence, not menus.
Decisioning logic adapts to traveler preferences, budgets, and itineraries.
Trust built in.
Secure, encrypted speech-to-text protects every customer interaction.
Production scale, before the AI wave.
Deployed across the Group’s global passenger base before generative AI became mainstream: voice-first engagement at the scale of one of the world’s largest carrier groups.
0M+
passengers annually, across the Group
One platform · five airlines
Live 3D · the Group’s network, drawing itself
A platform, not a feature.
Mobile, in-app, conversational web: every new channel inherits the same AI, accessibility, and security baseline without re-engineering. Per-channel rebuilds became platform-level reuse. Designed and deployed by Arslan Khan with Lufthansa’s Digital Hangar.
145M+
passengers a year served by one intelligent platform
5
airlines on the same AI, accessibility, and security baseline
2022
shipped, before generative AI became mainstream
Mercedes
An AI parts-counter rep that answers by Year/Make/Model, checks stock, and handles calls, live on the storefront with zero setup.
“2019 Ram 2500” is all it needs.
Live on the storefront · zero setupHey Mercedes, I need a rebuild kit for my 2019 Ram 2500.
Got it. 2019 Ram 2500 with the 6.7 Cummins, that’s a 68RFE. Our Category 0 kit is the right starting point, in stock and ready to ship today.
SC-68RFE-K0
68RFE Category 0 Rebuild Kit
✓Fits your 2019 Ram 2500
“Want me to add fluid, or are you good there?”
One message in: the agent’s trace
Automotive structure, not a generic chatbot: year, make, engine, transmission, catalog
Bundles, orders, payment, without leaving the chat.
“Smart. Here's a bundle that covers the rebuild, fluid for a full service, and the triple-disc billet converter. Saves you about $340.”
The Towing Rebuild Bundle
68RFE Category 0 Rebuild Kit
SunCoast Full Synthetic ATF · Case of 3
Guardian HD · triple-disc billet converter
Built on the fly, from one follow-up question
“Order SC-104287 shipped Tuesday, FedEx Ground, should be at your door Friday.”
Order SC-104287
● SHIPPED68RFE Category 0 Rebuild Kit · ×1
SunCoast Full Synthetic ATF · ×2
Pulled live from the store’s order system
“Done. Sent to the number on file ending in 4471. It expires in 24 hours.”
Payment link sent
TO ···· 4471 · EXPIRES IN 24H
pay.suncoastdiesel.com/t/8kq2…Checkout finished by text: no cart abandonment
Same conversation, your channel.
One thread, six doorsMercedes remembers the truck, the order, and the thread, wherever you pick it back up.
Web chat
“Hey Mercedes, I need a rebuild kit for my 2019 Ram 2500.”
In-chat voice
Phone
SMS
“Shipped. FedEx 7712 4451 0928, arriving Thu.”
Discord
#diesel-builds “@Mercedes which fluid for the 68RFE?”
A commerce agent that acts, not chats.
Fitment intelligence over real automotive structure, live catalog grounding, in-conversation checkout, production voice, and multi-channel memory. Engineered end to end by Arslan Khan on WSM, the platform built for automotive complexity.
1
message: year/make/model to verified fitment
6
channels, one continuous memory of the customer
24/7
on the counter: answers, bundles, orders, payment
Scrively
Generative AI that writes the story, illustrates it in 110+ art styles, narrates it, and ships it in print and digital.
One photo becomes the hero of the book.
✨ Story · art · voice in one chainChoose a character
✨PHOTO → PIXAR-QUALITY CHARACTER
Created once, saved to the library: Wilson, Draco, Bora star in every next story
STYLE: ANIMATED 3D · 1 OF 110+They set off on an adventure through the enchanted forest, where the trees whispered secrets and the flowers danced in the breeze.
Any of these, starring your child.
✨ Personalization takes about 60 seconds
✨ EVERY ONE OF THEM · GENERATED
Kids choose what happens next.
Choices, branching paths, and clickable elements turn the book into play. The reader becomes part of the story.
Scrively’s output is a real book on a real shelf.
Milo and the Moonlit Footprints
SQUARE 8.5″ · 31 PAGES
Ordered
One click from the reader · 1 credit
In print
Print-on-demand · hardcover case wrap
Shipped
Tracked, worldwide methods
On the shelf
A real book, with their child inside
Digital twin included: PDF reader on any device, a shareable link for the grandparents
→
One click more: a coloring-book edition. Same story, same scenes, kids color them in.
Generative AI, end to end: down to the bookshelf.
A multi-model pipeline orchestrated by Arslan Khan: story generation, character-consistent illustration in 110+ styles, narration with music and sound effects, branching interactivity, layout, and print-on-demand fulfilment. One production system.
~60s
from one photo to a finished, personalized book
3
generative modalities in one chain: story, image, voice
24h
to hardcover production: print-on-demand, integrated
Muzaura
AURA, the verdict engine for audio: five-signal authenticity scoring, cryptographically sealed into the file itself.
Sound in. Evidence out.
AURA Engine · queue-based pipelinevalidate · checksum
LUFS · clipping · silence
CLAP 512-dim
MFCC · YAMNet
five-signal score
sign · seal
What the workers extract · never the audio itself
Privacy-safe by design: audio becomes non-reconstructable feature vectors.
Models score the signal’s features, not a copy of the recording. GPU workers, 10-second windows, 50% overlap.
Five signals. One score it can defend.
SVI · Signal Verification IndexCORE 0.831 → γ 1.4 → 0.768 → SVI 77
AI-LIKENESS 0.22 · ORIGIN HUMAN
Component scores · weights from the scoring config
LUFS · true peak · clipping · silence
LightGBM on CLAP embeddings → AI probability
MFCC variance across 5-10s windows
Micro-timing stability · post-MVP, weight redistributes
FAISS similarity · duplicate detection, vault-scoped
When signals disagree, or AI probability sits in the uncertain band, it asks for context instead of guessing.
The verdict is sealed into the file itself.
Cryptographic provenanceAudio + XML + evidence pack, packaged and submitted. The AIID travels to every storefront.
Under the product: a real distributed system.
MVP in 17 weeks · 3-4 engineersProcessing flow
aura.logging
every event → audit trail
aura.notifications
async delivery
aura.etl
→ ClickHouse analytics
aura.dlq
dead letters · 7-day TTL
SCALE-FROM-ZERO AUTOSCALING
QUEUE DEPTH 70 · TARGET 5 MSG/TASK → CLOUDWATCH → ECS
WORKERS 0 → 0
Built and operated by Arslan Khan
COMPUTE
ECS Fargate · 10 worker services · GPU pool
QUEUES
RabbitMQ · durable · retry + DLQ policies
DATA
PostgreSQL · Redis · ClickHouse · FAISS
STORAGE
S3: raw, evidence packs, DDEX, JSONL logs
IAC
Terraform end to end · multi-env parity
COMPLIANCE
7-year audit retention · KMS-held keys
Dual-model signal interpretation: construction-profile + generator-alignment LightGBM on CLAP, calibrated outputs, with a data flywheel and monthly retraining as an operational commitment.
Trust infrastructure for audio: designed, built, shipped.
The AURA verdict engine, a cryptographic provenance layer, and the queue-based GPU pipeline underneath them, all architected and operated by Arslan Khan end to end, from the Terraform up to the monthly model-retraining cycle.
100+
files a minute, the pipeline's design target
5
signals behind every score, weighted and calibrated
1
cryptographic seal that travels with the file, verifiable anywhere
Seven engagements · six industries · one architect
Seven systems.All shipped.
Any world is one click away for questions
Arslan Khan · Systems Architect