Machines that show up in a tie.
terminalwork is an IT consulting and product studio. We move infrastructure to the cloud, put AI on the device instead of the server, and write the software that connects the two — then we ship it.
services
Three disciplines, one team. Most of our engagements pull from at least two of these at once — a cloud migration that ends with an edge model, or a backend rebuilt to feed one.
Software Development
- Backend systems and APIs built to hold production load.
- Cross-platform apps in Flutter and React, with native modules written by hand when the SDK runs out.
- Standalone packages and SDKs, designed to be published and maintained.
Cloud
- Migrating existing infrastructure to the cloud with minimal downtime.
- Auditing and cutting the cost on infrastructure that's already live.
- Deploying applications to production fast, with a pipeline that doesn't get in the developers' way afterward.
AI
- On-device and edge LLM deployment: TensorFlow Lite, custom tokenizers, zero cloud dependency if needed.
- Interpretability and bias auditing using activation patching and causal ablation.
- Fine-tuning and architecture guidance for teams moving from a cloud model to a local one.
products
Alongside client work, we build our own. Some of it ships as apps, some as open packages other engineers depend on.
Velar
Type what you ate, in plain language, and Velar extracts calories and macros in seconds — entirely on-device. No photo guessing, no food databases to scroll through, no data leaving your phone. Syncs with Apple Health, stays offline the whole time.
ask us about itSpot
Spot is an AI fashion discovery platform in Egypt. Users search using images or text to find similar local clothing. Built on serverless infrastructure, with VLMs to process catalogs and auto-label items in the background.
visit spot.terminalwork.netBERT Tokenizer
A pure Dart WordPiece tokenizer for BERT-style NLP models — no Flutter dependency, so it runs on backend servers, CLIs, the web, or inside an app. Handles subword splitting, padding, and attention masks, and outputs arrays ready for TFLite or ONNX. Now past 300 downloads a month.
view on pub.devhard problems we've solved
A sample of the work that didn't have a template to follow.
Circuit Localization in a Tiny LM
Published research mapping how a small language model routes geographic knowledge internally — and why that routing resolves later, and less reliably, for MENA geography than for Western geography. Past 100 downloads on Zenodo/CERN.
read the paper & codeCPU-Optimized Inference Pipeline
Cut serverless inference costs from ~$200/month to under $40/month on AWS Lambda with SnapStart, holding sub-5-second latency across a chained three-model pipeline.
Cloud IoT Feeding System
An MVP-scale AWS backend (EC2, RDS, S3, IoT Core) for hardware in the field — automated device provisioning, plus real-time WebSocket video ingestion that sidesteps MQTT's payload limits.
credentials
The people building these systems also study how they fail, which shows up in how we work.
Certified AWS Solutions Architect Associate, with production workloads shipped on both AWS and Google Cloud.
Worked at Deloitte's Innovation Hub based in Cairo, EG on an AI-assisted SOC triage pipeline that cut alert response from an industry-average 5–15 minutes down to 1 minute.
Delivered a symbolic solver project to Julia Labs, now merged into the Symbolics.jl library which is used by Julia developers worldwide and in industry products (1.5K stars on github).