Machine Learning Engineering.
Production ML: feature stores, serving, retraining, drift alarms. Not notebooks. We start from the eval and the cost model, ship what moves the metric, and alarm on drift before your users feel it.
What shows up in your repo.
Agent graph
Typed nodes, tool schemas, retry and fallback policy. Versioned like code.
agents/*.ts · dag.yamlEval harness
Golden sets, LLM-as-judge, regression runs on every PR.
evals/*.jsonl · ci.ymlObservability
Traces, token costs, hallucination rates, drift alarms. Wired to your stack.
otel · datadog · honeycombRunbook
What to do when a tool 500s, when latency spikes, when eval red-lines.
docs/runbook.mdLLM systems that earn their place in production.
Most teams reach for fine-tuning too early. We start from the eval and the cost model, then decide what actually moves the metric.
When to fine-tune
A clear-eyed call between prompting, RAG, and fine-tuning — based on your data, latency budget, and the behavior you actually need, not hype.
Data & evals
Dataset construction, golden sets, and LLM-as-judge graders so a fine-tune is measured against a baseline, not a vibe.
Serving & cost
Inference that holds latency and budget — quantization, caching, batching, and routing across hosted and open-weight models.
Guardrails
Safety, PII, and policy graders in the request path, with nightly regression runs in CI.
Arabic-language models, evaluated like the rest.
Less than a sliver of public LLM eval data is in Arabic. For teams serving KSA, the UAE, and Egypt, that gap is the work — and it is where we are deep.
AR/EN eval datasets
Bilingual golden sets generated from your production logs, so Arabic quality is measured, not assumed.
Dialect coverage
Beyond Modern Standard Arabic — Gulf and Egyptian dialect handling where your users actually write.
Arabic RAG
Retrieval, chunking, and re-ranking tuned for Arabic morphology, with citations and groundedness checks.
In-region deployment
Patterns for data residency and in-country hosting, built into the standard runbook for sovereign workloads.
Drift is a measured event, not a surprise.
Every model ships with a registry entry, a golden dataset, and nightly graders. When a metric moves, an alarm fires before a customer notices.
A pricing engine, rebuilt in six weeks.
One representative case. Read the full editorial walkthrough, including the deploy log and latency chart.
Rebuilding a pricing engine in six weeks.
Series C fintech · monolith → 4 services + gateway · 42 TB ledger migrated · zero downtime.
READ THE CASE →Six things finance + eng teams want answered.
Both. We build classical ML (feature stores, serving, retraining, drift monitoring) and LLM systems (prompting, RAG, fine-tuning, evals). The discipline is the same: measure first, ship what moves the metric, then operate it.
When you need a behavior or format that prompting and retrieval cannot reliably produce, you have the data to support it, and the eval shows a real lift over the baseline. Often the answer is RAG first; we will tell you honestly which one your problem needs.
Yes. Arabic-language work is a core capability — bilingual AR/EN eval datasets, dialect coverage, Arabic-tuned retrieval, and in-region deployment for data-residency requirements in the Gulf.
A model registry, golden datasets, and nightly LLM-as-judge and metric graders. Drift triggers an alarm and a triage runbook before it reaches users — it is a measured event, not a surprise.
We engineer for your budget: quantization, caching, batching, and routing across hosted (Claude, others) and open-weight models. We target sub-40ms for classical serving and hold token budgets for LLM workloads.
Yes — on your cloud account under your key management, including in-region and sovereign-cloud patterns when data residency is a requirement.