Jace Sabr — AI Engineer

Founder, Jace AI Solutions · Inference Engineering Expert · Working globally.

Consulting

Jace AI Solutions

My AI consulting and automation firm. I work one-on-one with startups, small businesses and governments, building AI that fits the way they already work — from a quick prototype for a five-person team to a national rollout — and training the people who'll actually use it. Every project starts and ends with me personally. Deepest experience is in hospitality: a voice concierge and a self check-in kiosk in production for a hotel in Germany, with the same work pitched live to properties in India.

Expertise

Technical Skills

  • Running LLMs in production — serving, speed and cost at scale
  • AI agents that do real work, with guardrails on what they may touch
  • Realtime voice agents that take live calls
  • Making software work with systems that have no API
  • Answering from a business's own documents and data
  • Fine-tuning and deploying custom models
  • Full-stack web apps, APIs and payments
  • Computer vision — cameras, detection, on-device
Ongoing Education

Books & Workshops

Continuous, hands-on study — keeping current with where production AI is actually going.

Inference Engineering Workshop
Completed — Vizuara AI Labs, run by Dr. Raj Dandekar (MIT PhD). How to serve large language models in production: making them fast, cheap and small enough to run on real hardware, including on-device.
CompletedInference
Writing a Book on Inference Engineering
In progress — 27 chapters on how large language models are actually served: the five numbers you end up living by, what a GPU looks like from inference's point of view, the memory the model keeps while it writes, and every technique that exists to shrink it. Written in plain words on purpose — the test for every page is whether someone who has never opened a machine learning paper can read it and explain it back.
In progress27 chapters
Showcase

Projects

Products in the hands of real users, and the research behind them.

Clippr — Finds the Clips Worth Posting
Hand it a film or an episode and it hands back the moments that stand on their own, cut as vertical shorts. It reads the whole script first to find them, then reads each one back cold — the way a stranger scrolling past would meet it, with no knowledge of the rest — and throws away anything that only made sense because you had watched the other two hours. The picture is never cropped to fit the tall frame: the full shot is kept and the frame is built around it, so a widescreen film arrives whole. A presenter can introduce each clip in under ten seconds. The page is small and the video work is heavy, so each run rents a real machine in the cloud for the minutes it needs and gives it back.
FlagshipVideo AILive
myGuru — A Companion That Remembers
A daily companion app that actually holds on to you — the facts of your life, the events you mentioned once, and the threads you opened and never came back to, which it can raise later. Each day opens on a wheel that lands on one area of your life, decided by real astrology maths worked out for you, at that minute, where you are. The rule the whole thing is built on: the AI never looks anything up and never does arithmetic — every planet, clock time and remembered fact put in front of it was worked out by our own code that same second, so its only job is choosing the words. Android build signed; in closed testing ahead of release.
AndroidLong-term memoryClosed testing
Kushvi — Wear It Before You Buy It
A clothing storefront where you give it one photo of yourself and then click through the entire catalogue wearing each shirt in turn, instead of taking one picture per item. The products, the photos and the prices are real, pulled from a working Indian storefront, and the layout was measured off that storefront rather than guessed at. Built deliberately small — plain pages and one little server whose only job is to hold the image model's key — so the try-on is the only moving part.
Virtual Try-OnStorefront
Hotel Self Check-In Kiosk
A lobby kiosk that runs on a single Android tablet. Walk-ins book a room on the spot; guests who already booked check themselves in; either way the machine hands over the key in whatever form the property already uses — an RFID card, a key-box PIN, or a door code. Behind the screen it works the hotel's existing booking system — which offers no way in for other software, so it signs in and operates it the way a receptionist would, writing a real reservation in about a minute. The guest pays by tapping a card on the tablet. Built for a hotel in Germany.
FlagshipClient WorkLittle HotelierPayments
My Personal Teacher — Maths Tutor
A maths tutor that finds the one thing a student actually misunderstood and fixes it from the ground up, instead of re-explaining the whole chapter. It teaches, checks the student can do it, then tests — and the check is a real marked answer, not the model's impression of how it went. Built for CBSE Class 10/12 and JEE.
FlagshipTutorMaths
First Hello — AI Application Agent
An agent that applies to jobs and apartments on your behalf. You give it your real history once; for each listing it finds the posting, writes a fresh application in that listing's own language, fills the form and sends it — so getting a foot in the door stops depending on how many hours you can spend typing the same details into different forms. It never invents a credential, and it never sends the same template twice. Won $20K in funding from E2B.
$20K E2B fundingAI Agent
DeskHand — Custom AI Agency
My delivery brand for custom AI builds: agents that answer the phone, watch camera feeds, run software for you, and get businesses off paper. What makes "it works with whatever software you already run" a real promise is an agent that looks at the screen and drives it — clicking and typing like a person — instead of a scraper wired to one site that breaks when the page changes. Proven by four real submissions into four different enterprise hiring systems.
LiveVision AgentAutomation
realput — Measuring Filler in AI Output
A measurement I designed for a gap in how the industry rates AI systems: every standard speed metric rewards a model for producing words quickly, and none of them ask whether the words were worth reading. realput grades a response piece by piece — useful, filler, or wrong — so a model that padded its answer scores worse than one that answered and stopped. The grading works; teaching a second model to do the labelling reliably is the part still open.
ResearchMy own metric
Voice Concierge — Hotel & Clinic Clients
A voice agent that answers the phone for a hotel. Guests speak normally, in German or English, and it answers about rooms and rates, takes bookings, and hands over to a human when it shouldn't be deciding alone. The rule it's built around: every price, date and room it says out loud is fetched live from the hotel's system while it's talking — it is not allowed to guess, because a confident wrong price is worse than no answer. In production for a paying hotel client in Germany, and running as a live pitch for a hotel and a clinic in India. The same agent also handles the website and email.
Client WorkRealtime Voice
My Personal Teacher — Interview Prep
The same tutor engine as the maths version, aimed at engineers preparing for interviews on how large language models are actually served in production. It starts from the pieces a candidate is expected to explain on a whiteboard and builds up one concept at a time, checking understanding before moving on rather than handing over a reading list.
TutorInterview Prep
Glow — Wellness Listings Canada
A national directory for spa, massage and beauty services in Canada, carrying over 3,500 listings. Every city-and-category combination gets its own page generated automatically, so the site ranks for the searches people actually type, and businesses can pay to post themselves. Built and shipped end to end — the directory, the pages, the search and the payments.
Full-Stack3,500+ listings
Morrigan — Synthetic Human
An AI companion that remembers past conversations, builds trust over time, has thoughts it doesn't say out loud, and disagrees with you when it thinks you're wrong — rather than agreeing with everything. Built on a model I fine-tuned myself.
FlagshipFine-tuned model
Education

2025 — Present

Advanced AI & Robotics

Itronix Solutions — AI/ML, deep learning, CV, NLP, robotics. Hands-on delivery with real projects.

2020

BSc Computer Science

University of British Columbia. Algorithms, systems design, software engineering.

Let's work together

AI strategy, implementation, or a technical second opinion.