Roshaan Ejaz — independent AI systems consultant.
Real estate AI systems architect
Nobody. That’s what I fix.
Last system I shipped: first deal closed in week one
Systems I build on
I work as your technical lead, not as another vendor.
That means I'm the one asking the vendor hard questions, not the one dodging them. I build what I recommend, I stay on to run it, and I hand over every credential and every document. If the honest answer is that you don't need software, you'll hear that on the first call.
Services
Take the full system, or the single layer that's holding back the rest. Most engagements start with one and grow into the third.
01 / CAPTURE
Answers the line at any hour, knows your inventory, qualifies against your criteria, and books the tour into a leasing agent's calendar.
02 / CONVERT
The unglamorous layer that decides whether leads become deals. Most outreach dies on the second and third touch — exactly the part a busy human stops doing.
03 / ADVISE
For groups already buying AI and ERP work. I sit on your side of the table, audit what the vendor delivers, and translate it for the people paying for it.
Track record
What was broken, what I built, and what changed. No logos I haven't earned and no numbers I can't defend.
Week 1
Time to first closed deal on an outreach system I built
159
Custom ERP modules audited as technical lead
2
Languages spoken by the voice agent, both natively
24/7
Coverage on an enquiry line that used to close at six
Every enquiry now gets a same-day conversation instead of a voicemail.
Enquiries arrived faster than a small team could call back, and callbacks came late. Buyers expect to be spoken to in Urdu, on the phone, quickly. Off-the-shelf products failed twice over: wrong language, and no legal route to place outbound calls to local numbers.
The hard problem was never the AI. It was calling compliance: which entity may lawfully place automated calls, where media is hosted, how consent is evidenced. That is why this runs while most pilots stop at the demo.
First deal closed in week one. That deal covered the entire build.
Messages went out in bursts when there was time, follow-ups depended on memory, and replies sat in a personal inbox. No way to tell which conversations were still alive. Small enough to feel manageable, large enough to leak deals.
The obvious fix — a bulk sender — is also the fastest way to get a business number banned. That constraint shaped the build.
Nothing here is clever. It is consistency, made automatic, on a channel where people reply, run so the number survives past month two. Most outreach dies on the second and third touch — exactly the part a human stops doing.
159 modules audited. Migration risk quantified before go-live, not after.
A group migrating eight years of custom ERP work with an external vendor. Nobody inside could tell whether what came back would match what finance relies on daily. The people best placed to judge could not read a codebase, and the people writing it were reporting on their own progress.
I don't write the migration. I'm the client's technical counterweight to the vendor, and I own one question: will this behave the way the business needs on day one.
An AI fraud detection layer over the group's approval flows, scoped as an external service returning approve, reject or uncertain into the existing steps — testable and reversible without touching finance operations.
How I engage
Plenty of clients only need the first, and that's a fine outcome.
STAGE 01
I map where enquiries stall today and what your existing tools already do that nobody uses. You get a written architecture and a costed recommendation — including the honest version where the answer is a process change, not software. Yours whether or not you build with me.
STAGE 02
Built, integrated with your CRM and numbers, and put in front of real leads in a controlled slice before it takes full volume. You get the documentation and the credentials. Nothing runs on an account you can't access.
STAGE 03
Live systems drift. Prompts need tuning against real calls, templates get rejected, providers change the rules. I hold the system, report on what it produces, and act as your technical lead with any other vendors.
Working stack
I'm not tied to a vendor, and I'll tell you when the cheaper tool is the right one.
About
I'm Roshaan Ejaz. I build AI systems for companies that sell and manage property, and I stay on to run them.
I came to this through engineering, not marketing: a computer science degree, then years building automation and machine learning systems in-house before going independent. That's why the work survives contact with production. I care about the parts nobody demos — what happens when a provider rate-limits you at 4pm, whether your calling setup is legal where you're calling, and who owns the account when the engagement ends.
Property is a business of fast follow-up in more than one language, which is exactly what these systems are good at. I work with a small number of clients at a time, across the US, Europe and Pakistan.
Earlier engineering work: a retrieval system over clinical documents using LangChain and FAISS; transformer models for code generation and summarisation in PyTorch; an automated unit-test generator for Java evaluated across 50+ programs, built as a final-year research project. Not real estate, but it's where the retrieval and evaluation habits came from.
Writing
The telephony rules, the failure modes, the reliability work nobody demos. Written up as I hit them.
FAQ
Products are built for the average operator. If your stack is standard and your process is standard, buy the product, and I'll tell you which one. I'm useful when the pieces don't fit: a second language the product doesn't really speak, a CRM nobody integrates with, telephony rules in your country, or three tools that each hold part of the lead record. Increasingly the work is fixing AI that's already been bought and isn't landing.
The diagnostic is a fixed fee. Builds are fixed scope, quoted after the diagnostic so we're both pricing something real rather than a guess. Ongoing support is a monthly retainer, and advisory engagements are retained by the month. No hourly billing, because you shouldn't pay more when something takes me longer than it should.
A working voice agent on a controlled slice of traffic usually takes four to six weeks from signed scope. Outreach systems are faster. The long pole is almost never the AI. It's phone number provisioning, WhatsApp template approval, and getting access to your CRM, so the sooner those start, the sooner you're live.
You keep everything. The systems run on your accounts, with your API keys, under your billing, from day one. You get the architecture documentation and a written handover. If you want to move it in-house or to another vendor, nothing about my setup makes that hard, which is deliberate.
I work US and European hours and I'm reachable during your day. More to the point, I take a small number of clients at a time specifically so that response time is short. If a live system breaks at your 9am, that's my problem, not something that waits for a ticket queue.
Yes. References are available once we've had a first conversation and I know which engagement is the relevant comparison.
Start here
Roughly how many enquiries you get a month, and what happens to the ones nobody calls back. Thirty minutes. If I'm not the right person for it, I'll say so on the call and point you somewhere better.