Head of AI/ML · Dubai, UAE
Two years of experience as an engineer, and a master's in Analytics & AI from Imperial College London, ranked second in the world. I like to build distributed, scalable systems. Most recently an AI Implementation Strategist at RBC, researching how agentic AI tooling integrates into global equities divisions.
01 — Education
MSc Business Analytics & AI · First Class Honours
AI / Machine Learning
Mathematics
Data & Systems
Dissertation on a contextual multi-armed bandit for promotional offers, below.
BSc Information Management · First Class Honours
Analytics
Data & Systems
Programming
Fig. IThe dissertation's method: Thompson sampling on three offers over 240 pulls. The readout is measured off the history.
02 — Work
AI Implementation Strategist
Researching how AI and LLM tooling integrates into investment research for the global equities team: which steps of the process they can take, and how to check the result.
Founding Engineer
An online grocery retail delivery platform, areednow.com, delivering FMCG at scale.
Machine Learning Engineer
Data Engineer
An automated sales-data pipeline from harvesting through cleaning, aggregation, streaming and storage; then customer analysis and clustering on top of it to find market cohorts the business was not yet selling to.
Recruit · United Arab Emirates
Ranked first of a national service cohort of more than 1,000 soldiers.
Fig. II aOrder allocation across the fleet: 48 orders, 4 vans of 13, assigned by regret and re-centred until the drive stops falling.
Fig. II bThe Areednow platform as built: six clients on one Postgres, the edge functions and outbox around it, the models it calls, and the analytics plane fed by change capture. Hover a layer.
Areednow v2 — one monorepo, six front ends, one Postgres
Browse, cart, Stripe checkout, live order tracking over Realtime.
Route stops, a location ping every few seconds, messages with the customer.
A live picking queue, packing, dispatch; the floor's only screen.
Catalogue, inventory, purchase orders, returns and rebates, dispatch.
Order messages and WhatsApp through Twilio, in one queue.
Marts from the lakehouse, so analysis never touches the live database.
Before it, and beside it
The first customer app: push through FCM, KNN semantic ordering of the catalogue.
The first back office: inbounds, purchase orders, rebates, barcode scanning on the floor.
Competitor prices scraped on a schedule, signals to Slack, a dashboard over its own Postgres.
Fig. II cThe applications, each with who uses it and what it runs on. Six ship from one monorepo; the two v1 apps and the price monitor stand apart.
03 — Projects
Four projects, numbered newest first and each built by me alone, end to end. Each says what it was and what problem it solved, and carries a figure that computes the method it used.
A rules chatbot for Jiu Jitsu competitors, coaches and referees.
I built a retrieval-augmented chatbot over the IBJJF, ADCC and AJP rulebooks, with a fine-tuned retriever, a semantic cache and an LLM judge validated against human labels. It solves the problem of three rulebooks that disagree and change every season, and of chatbots that answer without a source: every answer cites the section it came from and the hallucination rate is measured.
III aTop-k retrieval by cosine on a ring of chunk embeddings. The arcs are the semantic cache at the τ set below.
A spot market for GPU compute and inference capacity.
I built a decentralised exchange where owners of idle GPUs post asks, workloads post bids, and the market clears at the crossing. It solves the mismatch of GPUs sitting idle on long contracts while small inference jobs queue or overpay for on-demand instances.
III bEight jobs to six GPUs: a greedy first-fit pass, then the optimal assignment by the Hungarian method. Both costs measured.
A crop disease prevention and detection system for farmers.
I used ML, forecasting models, optimisation and network diffusion models to detect crop diseases from images, estimate their spread to other farmers, and forecast medicine requirements and the shortest delivery times. It solves the three problems a farmer has at once: what the disease is, whether the treatment is in stock, and whether the neighbouring farms are next.
III cAn SIR epidemic over 50 farms and their 4 nearest neighbours. Infected in the accent, recovered filled.
A review generator, and a detector that catches it.
I built a from-scratch BPE tokeniser, an LSTM language model that writes Amazon-style reviews, and a 1D-CNN discriminator that detects them at 91% accuracy. It answers the two questions a review platform cares about: how convincing a small model's fakes can get, and whether they can still be caught.
III dByte-pair encoding on a review corpus. The most frequent pair merges each step; a sentence retokenises.
04 — Research
MSc dissertation at Imperial: a contextual multi-armed bandit for promotional offerings that chooses the product, the discount and the duration together, built on Jain et al. (2023) and Ban and Keskin (2020). The same system is going into production at Areednow, where the context is a live grocery basket and the reward is margin, not a simulated click.
Fig. IVA contextual bandit: the best offer flips with basket size, and one posterior per offer and context learns it. Regret against a context-blind bandit on the same arrivals.
05 — Skills
This is the tech stack I use regularly, sorted by engineering role and by the layer of the system where each tool runs. The columns are the three roles: software, AI/ML and data engineering. The rows are the layers of a production system, from the client down to the infrastructure. Read down a column for the reach of a role, and across a row for what runs at a layer.
Fig. V32 tools across three disciplines and eight layers. Each is placed once, where it was used most. Hover a cell to read its lane and band.
06 — Languages
Spoken
Programming
07 — Hobbies
Three of them, and each one is a search problem: a position, a set of legal moves, and a cost of getting it wrong. Each gets the figure it earned.
01 — Jiu Jitsu
Blue belt, competing at −70 kg since 2021. The rulebooks I trained a chatbot on are the ones I compete under.
02 — Chess
Peak 1314 on chess.com. Ruy Lopez as White, Caro-Kann as Black. The board beside plays both from the move list, and turns round for the second.
03 — Rubik's cube
A personal best to the hundredth. Forty-three quintillion positions, and every one of them is at most twenty moves from solved.