RAG systems
I design and tune retrieval pipelines end-to-end: chunking, embeddings, hybrid search, and reranking.
I'm Salim Jordan,
an AI Engineer building
RAG systems that work.
I build, measure what matters, and share what I learn. No "thought-leadership speak", just findings with numbers.
What I do
Evaluation pipelines
I build measurement workflows with retrieval metrics like MRR, NDCG, and Recall@K to prove what actually improved.
LLM products
I ship LLM-powered features with structured outputs, tool use, and practical reliability constraints.
Who I am
I spent 25 years in enterprise consulting working with companies like MarkLogic, Oracle, and RightNow. I've led technical delivery for banks, telcos, and government agencies across APAC and North America. Now I'm applying that lens to AI engineering: evaluation pipelines, RAG systems, and the infrastructure that tells you whether an LLM actually works before it ships.
Background: Technical leadership and delivery where things need to work reliably, at scale, for real users.
Focus: Measurable outcomes over demos. Retrieval metrics, failure-mode analysis, production-minded AI.
How I work
1
Define success
What metric matters? What's the target? What does "good enough" look like?
2
Measure the baseline
Before optimizing, know where you are. Can't claim improvement without a starting point.
3
Iterate with evidence
Each experiment answers a question. The data tells you what to try next.
What I've built
Customer Sentiment & PM Intelligence
Four-agent pipeline scoring sentiment, discovering themes, and aligning to a roadmap across 4,742 reviews from Amazon, Yelp, and the App Store.
Accuracy: 84.5% Reviews: 4,742
PaperSearch
RAG system for 1,000 academic papers. Hybrid retrieval, cross-encoder reranking, validated against 3,045 human-authored queries.
MRR: 0.789 NDCG: 0.797
Recent findings
The Eval Said the Model Wasn't Needed Yet
I built the eval and a grep baseline before the model. The baseline tied the model on every real case, and the eval found two bugs I'd have blamed on the LLM.
Title-match baseline, mention recall
1.00
Claude Haiku 5.5, mention recall
1.00
Prompt v2 on the local model
-11pp recall