Major applied-AI case study
AIRIS
Applied-AI engineering experimentsApplied-AI engineering experiments measuring retrieval and LLM behavior, reducing unnecessary context, and tracing concurrency bottlenecks.
Summary
AIRIS is an enterprise retrieval-augmented-generation workload at the Aerospace Malaysia Innovation Centre. I worked primarily on applied-AI engineering experiments: measuring retrieval and LLM behavior, reducing unnecessary context, and tracing concurrency bottlenecks rather than changing the system based on intuition.
Context
AI-assisted analysis workflows combine retrieval, model calls, and application services. A successful single request does not explain whether context is necessary or which subsystem dominates under concurrent demand, so I separated controlled observations from later recommendations.
Controlled context experiment
AIRIS · Controlled experiment
Reduce context, retain checks
Read this visual as text
A three-stage AIRIS controlled experiment shows six scenarios and ten turns, local context reduced from 541k to 20.9k tokens by 96.1%, and 10/10 citation checks passed. Citation checks do not prove universal answer quality.
- The controlled experiment covered 6 scenarios and 10 turns.
- Local context fell from 541k to 20.9k tokens, a 96.1% reduction.
- All 10 citation checks passed; that result does not prove universal answer quality.
I designed a controlled RAG comparison to measure whether large amounts of locally supplied context were necessary. Across a 6-scenario/10-turn controlled experiment, the optimized approach reduced local context from 541k to 20.9k local-context tokens, a 96.1% reduction, while retaining 10/10 citation checks passed.
Passing the experiment’s citation checks does not prove that answer quality universally improved. The checks establish the result for the defined scenarios and turns; they do not support a general accuracy or production-quality claim.
Concurrency and bottleneck diagnosis
AIRIS · Bottleneck diagnosis
Vector search dominated request time
Read this visual as text
A three-anchor AIRIS diagnosis shows vector-search latency at 4.8 seconds with 10 concurrent users, 51.0 seconds with 50 concurrent users, and vector search at 76.9% of request time at 50 users. Outbound-LLM behavior surfaced as another optimization area.
- Vector-search latency was 4.8 seconds at 10 concurrent users.
- Vector-search latency was 51.0 seconds at 50 concurrent users.
- At 50 users, vector search represented 76.9% of request time; outbound-LLM behavior was another optimization area.
Instead of treating increasing end-to-end latency as a generic AI-performance problem, I decomposed request time by subsystem. Vector-search latency grew from 4.8 seconds at 10 concurrent users to 51.0 seconds at 50 concurrent users; at 50 users, vector search represented 76.9% of request time. The testing also surfaced outbound-LLM behavior as another optimization area.
These figures describe the observed test conditions, not a supported production concurrency level or a measured production improvement.
Engineering outcome and attribution
AIRIS · Engineering handoff
From diagnosis to recommendation
Read this visual as text
A compact three-stage AIRIS path shows controlled workload testing, bottleneck diagnosis, and an engineering handoff with recommendations accepted by the senior engineer and stakeholder training as outcomes. Several recommendations were later incorporated into the senior engineer's implementation; no production improvement is claimed.
- Hydar performed controlled workload testing.
- Observed behavior informed bottleneck diagnosis and optimization research.
- The findings became recommendations accepted by the senior engineer and supported stakeholder training; several were later incorporated into the senior engineer's implementation, with no production improvement claimed.
I presented the findings and remediation options to the senior engineer. The recommendations were reviewed and accepted during engineering discussions, and several were subsequently incorporated into the senior engineer’s implementation. I authored the testing and optimization analysis and supported stakeholder training; later implementation ownership remains explicitly attributed to the senior engineer.
Scope and limitations
This case study does not claim ownership of the overall AIRIS architecture, implementation of every recommendation, a universal answer-quality improvement, supported production concurrency, or a resulting production-performance gain. It omits private documents and data, internal service or endpoint details, model and configuration details, dashboards, report figures, credentials, and copied professional artifacts.