Building AI-powered products that ship.

Currently building AI video canvas & Copilot

AI Video & Agent Product Engineer with 10+ years of full-stack experience. Currently building an AI video canvas workflow engine and Copilot; previously led AI platform engineering at Creatify.ai. Open-source contributor (ArkID, 1.3k+ stars).

Read the Blog About Me

132+

Articles

10+

Years

1.3K

Stars

Latest Articles

View all →

2026.10.16AI

Stuffing a 24B Always-Thinking MoE into 24GB of VRAM: A Three-Batch Field Test of LFM2.5-24B-A2B

LFM2.5-24B-A2B, released by LiquidAI on 2026-10-12, is a 24B-total / 2.45B-active MoE that fits entirely into a single RTX 4090 at Q4_K_M, sustaining 224–257 tok/s in our tests. Across three batches of 11 production-code case studies, it aced tool orchestration and gating semantics, stumbled on a classic off-by-one in fine-grained implementation, and clearly trailed cloud flagships on long-form Chinese creative writing. A strong local Copilot; do not ask it to write prose.

2026.10.15AI

Scout Under Test: A Two-Day-Old Browser Agent Meets a Real Domain Hijack

Two days after Outcrawl open-sourced their browser agent Scout, we ran it through four benchmark tasks and three real-world research missions against two other driving harnesses. All three returned identical verified data; Scout ran 3-8x slower, and it was the only one that noticed the brand's official domain had been hijacked.

2026.10.10AI

The Endgame for Repository-Scale Code Cognition: A Deep Dive into AOCI and Enterprise Agent Engineering

When autonomous coding agents tackle codebases with hundreds of thousands of lines, they consistently fail from blind file discovery and unstated architectural invariants. This article breaks down AOCI-CODE, its Volumes v1 architecture, the FRAS semantic model, and how enterprise engineering teams navigate scope governance and concurrent worktree conflicts.

2026.10.08AI

Teaching the Harness to Improve Itself: Lilian Weng's Roadmap, Checked Against Our Practice

Lilian Weng compresses the near-term path of recursive self-improvement into a five-step ladder: the object being optimized moves outward from prompts to structured context, workflows, harness code, and finally optimizer code. This piece unpacks the key numbers across three research lines and checks each step against agent systems we run daily: execution records on disk, a system prompt diet from 106k down to 78k tokens, an audit of 113 skills, and why evaluators must stay outside the evolution loop.