Engineering Insights & Stories
Practical knowledge, architecture deep-dives, and startup strategy from the engineering team at Lobhari Technologies.
CybersecurityGLM-5.3 & The Rise of Cyber-Capable AI: Automated Vulnerability Discovery & Defensive Architecture
An architectural deep dive into Zhipu AI's GLM-5.3: a 743B open-weight foundation model scoring 84.5% on CyberGym, automating zero-day vulnerability discovery, and reshaping enterprise DevSecOps.
Reinforcement LearningInside Miles v0.1: The Production-Grade RL Post-Training Architecture for Trillion-Parameter MoE Models
An architectural deep dive into Miles v0.1 by LMSYS and RadixArk: scaling asynchronous RL post-training for 700B+ MoE models using SGLang rollouts, Rollout Routing Replay (R3), and Megatron-LM.
DeepSeekDeepSeek V4-Pro GA & V4-Flash: Inside the 1.6T MoE Architecture, 384K Output Limits & Agent Economics
An architectural deep dive into DeepSeek V4-Pro GA (1.6T MoE) and V4-Flash: 384K output token limits, 87.9% Terminal-Bench scores, and peak/off-peak agent pricing economics.
Thinking MachinesInside Thinking Machines Inkling: Mira Murati's 975B Open-Weight MoE Foundation Model
An architectural deep dive into Thinking Machines Lab's Inkling: a 975B sparse Mixture-of-Experts foundation model activating 41B parameters per query, engineered by Mira Murati's team for enterprise open-weight autonomy.
RAGBeyond Naive RAG: Why Cross-Encoder Reranking & Hybrid Search Beat 1M-Token Context Stuffing in 2026
An architectural guide on production RAG in 2026: why 3-stage hybrid retrieval with cross-encoder rerankers outperforms brute-force 1M-token context stuffing in accuracy, latency, and cost.
ValkeyValkey vs. Redis 8: The 2026 In-Memory Database Benchmark & Architectural Guide
An in-depth engineering benchmark and architectural comparison of Valkey and Redis 8 in 2026: multi-threaded engine performance, RESP3 wire compatibility, and cloud migration strategies.