Rakibul Islam
Contact
Featured Essay · #01 AI Perspectives Oct 28, 2026

Beyond Accuracy: Why I Think Reliable AI Matters More Than Ever

A model can perform exceptionally well on a benchmark and still struggle when the real-world environment shifts. Why future AI systems must understand the limitations of their predictions through uncertainty estimation, calibration, and adaptive inference.

RI

Rakibul Islam

AI Engineer & Researcher · 6 min read

Read Essay #01
uncertainty_routing.py
Evidential Reliability 
def adaptive_predict(x: torch.Tensor, tau: float = 0.35):
    # 1. Fast lightweight pass + Dirichlet evidence alpha = e + 1
    evidence = F.softplus(lightweight_model(x))
    alpha = evidence + 1.0
    S = alpha.sum(dim=-1, keepdim=True)
    uncertainty = alpha.shape[-1] / S  # Vacuity u(x) = K / S

    # 2. Escalate to specialist model when confidence is insufficient
    if uncertainty.item() > tau:
        return specialist_model(x), "escalated"
    return alpha / S, "accepted"
#AI Reliability · #Uncertainty · #Robustness Explore Full Essay

All Technical Essays & Reflections

Showing 10 of 10 articles
03 LLM Engineering
2026-10-16

RAG Is More Than Connecting an LLM to a Vector Database

Tutorials make RAG look like three lines of glue code between a PDF loader, a vector store, and an LLM. In practice, every stage—segmentation, hybrid retrieval, reranking, and faithfulness evaluation—determines whether the system is trustworthy.

#RAG #LLM #Information Retrieval
7 min read Read Essay
05 Technical Essays
2026-10-04

The Hidden Cost of Making AI Models More Accurate

A heavier architecture or multi-model ensemble might add 1.2% accuracy on a benchmark, while tripling latency and power consumption. Why adaptive, uncertainty-guided computation is a smarter path forward.

#Model Efficiency #AI Optimization #Inference
6 min read Read Essay