Work

Selected projects.

Products we have built and run for startups and businesses in India, Saudi Arabia and the UAE, plus our own AI platform.

AllMobileAI / RAGAWS

India · Flutter · AWS ECS · Redis · Kafka

100K concurrent requests at peak

Saudi Arabia · Flutter · Apps · APIs

Shipped to App Store and Play Store in 6 weeks

Product · RAG · LangGraph

Live with coaching institutes and education businesses

Dubai · Flutter · Payments

Real estate listings, released to both stores

AI capability briefs

Anonymised patterns.

Representative engagement patterns. Not client case studies. No customer is named, and no result figure is claimed.

Anonymised pattern

Problem pattern: Questions arrive on WhatsApp and Slack. The answers live in internal documents, plus a few operational tools.
Architecture: Channel adapters in front of a LangGraph router. Retrieval over the corpus. Tool calls only for actions the policy allows. A person can take the turn.
Our approach: Write the tasks the agent may take. Check retrieval before the agent is trusted with them. Keep a tool allowlist. Interrupt when a step would change a customer record.
Stack: LangGraph, RAG (chunking, embeddings, hybrid search, reranking), WhatsApp, Slack.

Anonymised pattern

Problem pattern: A domain task where prompting and retrieval still miss the format or the vocabulary the job needs.
Architecture: Approved examples become a dataset. A LoRA or PEFT adaptation is trained. An evaluation harness scores a held-out set. The adapted model is served beside the base model.
Our approach: Write the evaluation before training. Keep the base model as the fallback. Train only the adaptation, and only after the eval set shows the base model is the limit.
Stack: Open weights, LoRA / PEFT, an evaluation harness, AWS for training and serving.

Anonymised pattern

Problem pattern: Inference spend and idle capacity, not model quality.
Architecture: An inference service on AWS that scales with demand. A cache in front of repeated prompts. Batch work kept off the online path.
Our approach: Separate online and batch. Scale with the traffic you actually have. Cache prompts that do not change. Watch failures and token use so cost is visible.
Stack: AWS GPU inference, autoscaling, a cache.

Start a project

Engagements are fixed-price packages, shaped to the business. Book a call for a quote.