Stories by Abhijit (19)
Benchmarking KV cache memory management, PagedAttention vs RadixAttention prefix caching, and Ollama runner locking under 1 to 128 concurrent streams.
Inside the biggest architectural change to Go's standard library in a decade: SIMD-accelerated metadata probing, open-addressing flat layouts, and the production benchmarks you need to know.
How adversarial metadata in Model Context Protocol tools hijacks autonomous agent planners, why high-reasoning models are most vulnerable, and the production gateway patterns needed to stop arbitrary tool execution.
Deconstructing the WPC 5925–6425 MHz de-licensing notification, Wi-Fi 7 vs Wi-Fi 6 performance, and why concrete walls challenge theoretical 320 MHz throughput.
A direct lab evaluation comparing Apple Silicon unified memory throughput against AMD Strix Point and Hawk Point mini PCs for local LLM inference, quantized model capacity, and sustained thermal throttling.
How barrel files silently bloat First Load JS, why React 19 Server Components fail to tree-shake third-party libraries, and the exact Webpack/Turbopack configurations to pass Google INP and LCP.
A physics-level breakdown of optical flow interpolation, frame pacing queues, display scan-out timing, and the measurable input lag penalty of DLSS 3, FSR 3, and AFMF.
An architectural teardown of Microsoft's developer filesystem: Copy-on-Write block cloning, Defender asynchronous filter bypass, VHDX overheads, and the real-world benchmarks across npm, Cargo, and MSBuild.
Moving beyond legal memos: a technical blueprint for the statutory November 2026 Consent Manager interface, token revocation webhooks, and purpose-bound database schemas under India's data protection regime.
Separating corporate press releases from grid reality: NRC licensing timelines, HALEU enrichment bottlenecks, and why baseload nuclear reactors struggle with AI inference power spikes.
A transparent operational audit of Supabase 7-day pauses, AWS IPv4 and NAT gateway surcharges, Vercel commercial usage restrictions, and the exact recipes to host side projects at truly zero cost.
An engineering evaluation of V8 13.x default type-stripping execution, native SQLite, JavaScriptCore memory pressure, and N-API crash boundaries under sustained enterprise loads.
Deconstructing the WPC 5925–6425 MHz de-licensing notification, Wi-Fi 7 vs Wi-Fi 6 performance, and why concrete walls challenge theoretical 320 MHz throughput.
How adversarial metadata in Model Context Protocol tools hijacks autonomous agent planners, why high-reasoning models are most vulnerable, and the production gateway patterns needed to stop arbitrary tool execution.
A direct lab evaluation comparing Apple Silicon unified memory throughput against AMD Strix Point and Hawk Point mini PCs for local LLM inference, quantized model capacity, and sustained thermal throttling.
A no-fluff purchasing guide evaluating thermal dissipation limits, soldered RAM traps, display ergonomics, and real Linux compatibility for software engineers across three Indian price tiers.
A physics-level breakdown of optical flow interpolation, frame pacing queues, display scan-out timing, and the measurable input lag penalty of DLSS 3, FSR 3, and AFMF.
Moving beyond perimeter VPNs: an engineering guide to cryptographic workload identity (SVIDs), mutual TLS service-to-service communication, and Envoy-based Policy Enforcement Points.
Deconstructing the ₹76,000 crore India Semiconductor Mission (ISM): Tata Electronics 28nm Dholera fab, Micron OSAT in Sanand, cleanroom water logistics, and the global mature node race.