Aramco Project Planning AI
An SME-grade ambiguity and scope-creep analyser for dense Arabic project plans — deterministic engineering plus a fine-tuned LLM.
Saudi Aramco · Planning-document intelligence
Overview
Aramco wanted a system that behaved like a senior SME — read a dense project plan, flag every ambiguity and scope creep, and make the document as close to what gets built as possible.
This was when open-source LLMs were rare and poor at producing reliable structured data. Microsoft hadn't given them the bespoke thing they wanted, and they were evaluating teams. The brand, and the clock, made the pressure real.
01 — Architecture
Deterministic first, model only where needed
I led the requirement gathering and the architecture. The principle: solve everything we could with deterministic engineering, and lean on a model only where judgement genuinely mattered. The system had four moves — Gather the document, Plan the analysis, Add Knowledge (Aramco's standards and references), and RAG over all of it, retrieved with a multilingual embedding model (BGE-M3) suited to Arabic.
02 — The model
Choosing a brain we could run ourselves
Open-source LLMs were thin on the ground and weak at structure. We evaluated the field — largely on hard-won intuition — and landed on Nous-Hermes: the most reliable instruction-follower we could host and control ourselves.
- Llama 2
- Mistral
- Mixtral
- Zephyr
- Nous-Hermes
03 — Structure
50+ keys, every time, no hallucination
Even the best base model wouldn't reliably emit what we needed — one response carrying 50+ keyed JSON fields, valid and complete, with no hallucination, inside a 128k context where about 50k was the safe working zone. So we fine-tuned it on a large body of synthetic data until it followed instructions exactly and produced the full structure every time.
Proven on real projects
The whole system shipped as a working platform POC — and we put it on real work. Each project was audited with the tool and signed off by top management.
- Underwater oil pipelines
- Solar parking lot
- …and more
Catch the ambiguity on paper — not on the seabed.
Role
Led requirement gathering and architecture for the engagement — the deterministic-first design, the model evaluation and fine-tuning, and the RAG system — delivered as a working platform POC under real time pressure, on-site.