TLI Tech

TLI Tech Blog

Engineering notes from AI systems built for real production.

Focused writing on AI Agents, LLMs, Speech AI, and infrastructure optimization. Clear engineering decisions, repeatable workflows, and lessons that can be applied.

Latest writing

04
Playful four-stage abstract organism illustrating how an LLM processes a request

AI Agent / LLM · Deployment & Optimization · 5 min read

What happens when an LLM answers a request?

A concise guide to prefill, the first output token, KV cache, and the decode loop behind every LLM response.

Sep 03, 2026
An eagle guiding a document-scanning courier boat toward an archive

AI Agent / LLM · Deployment & Optimization · 5 min read

Faster enterprise OCR with EAGLE 3 and Qwen3 VL

A real document workflow using EAGLE 3 to accelerate Qwen3 VL inference by 2.5 times with nearly unchanged accuracy.

Sep 03, 2026
A tactile collage of tangled voice ribbons woven into one fast, precise output stream

Speech AI · Deployment & Optimization · 10 min read

Accelerating OmniVoice with a custom TensorRT LLM FP8 engine

Why TensorRT LLM was the right layer above TensorRT, how we converted the Qwen3 backbone to FP8, and what actually improved on an NVIDIA L4.

Sep 02, 2026
Minimalist performance trace highlighting a single hidden inference bottleneck

Speech AI · Deployment & Optimization · 7 min read

Profile first: finding the right path to faster AI inference

How profiling changed our understanding of an OmniVoice bottleneck and guided a practical optimization from 0.145 to 0.056 RTF.

Aug 25, 2026