blog

DeepInfer is DeepSeek's open-source inference engine for V4.1-Flash. A code-level tour of DSpark, expert parallelism, async scheduling and kernels.

Nemotron 3.5 ASR and Transcribe Lite on an M3 Pro via Core ML vs Orukeet and Parakeet: WER, speed, chunk tiers, empty output, first-word loss.

An unauthenticated RCE in React Server Components (GHSA-xqc2-xf4c-x3h3, CVSS 9.5) hits Next.js 16 and react-server-dom-* via typed-array Flight payloads.

NVIDIA Nemotron 3.5 ASR streaming explained: mel frames, cache-aware FastConformer, RNN-T, language prompts, chunk tiers, the 30B Transcribe, and code.

Speculative listening decodes the next audio chunk before it arrives: 2.6-2.8x faster, lossless. Why it fails for ASR, and where it works in dictation.

CVE-2026-123158 (sshole) lets an oversized password bypass OpenSSH on musl Linux without PAM. How to check Alpine, OpenWrt and Docker, patch, and harden.

How GitHub tokens leaked into AI gateway logs let attackers ship a credential-stealing LiteLLM 1.107.0 to PyPI and Docker, and what to rotate now.

Eight eMudhra (emSign) intermediate CA keys were tiny integers anyone could recover. Full timeline, CVEs, patched versions, and how to stay protected.

PHP2Shell (CVE-2026-109618/109619) is an unauthenticated RCE hitting PHP 5, 7 and 8. The root cause, the live exploitation, and how to patch and harden.

Chat with a TradingView chart: TypeSafe's Jev routes asks into typed commands, GPT-5 plans the hard ones, and Pine runs in a Web Worker on my own interpreter.

How I am building nalgo: pure-Rust algo trading on NautilusTrader for Indian brokers. Connect your broker, write a strategy in plain English, let it trade.

Profiling Parakeet on Core ML found 20 ms zeroing a buffer one NSNumber at a time. FluidAudio#941: 72.6 to 48.2 ms, and review made the fix smaller and safer.

TapTalk: Orukeet (oruk.ai) as default engine, audio and Silero VAD moved from Rust to Swift and Core ML, key-up to paste 155 to 100 ms, three mic bugs fixed.

How I rebuilt TapTalk's engine: NVIDIA Parakeet and Nemotron on the Neural Engine, streaming live typing, Hindi and Hinglish, and Smart Mode rewrites.

Swapping vLLM's suffix drafter without a fork: two gates in two processes, one import-binding mistake, one process-boundary mistake, and the plugin fix.

How I built slate, a free open-source phone-as-Stream-Deck for Mac over your LAN with no cloud: one Command, a pluggable Transport, one shared Zod schema.

Team SSH to a private Azure dev cluster with zero open ports: one outbound cloudflared bastion, Cloudflare WARP, Google SSO, and the split-tunnel routing trap.

Pushing Qwen2.5-0.5B to 9,052 tok/s on a free Colab T4 in the H2Loop contest: AWQ INT4 dead ends, FP16, and model-free suffix speculative decoding.

How I built Logpipe, a small but real Vector + ClickHouse logging pipeline in Go: why aggregator batching matters, 'Too many parts', and disk buffers.

How I built TapTalk, a free local speech-to-text app for macOS in Rust and SwiftUI: Whisper on the Neural Engine via Core ML, chip-tuned, push-to-talk.

How Swiggy, Airbnb, PhonePe and Nubank ship UI without app store review, and how I built a BDUI + OTA system in React Native with Cloudflare Workers and EAS.

Dissecting the April 2026 Vercel breach: infostealer malware to OAuth hijack to customer data exposure. Full kill chain, trust lessons, survival guide.

How we built a production fair-value calculator on top of Dhan - Black-Scholes, cost of carry, live depth, and mispricing signals - for Saras and NSE F&O.

How we built end-to-end automated prop trading at Saras: 1000+ daily trade ideas to live execution on Dhan, with novel metrics that actually generate profits.

How we built Saras's agentic news engine: scrape public macro and trade intelligence, dedupe at scale, enrich with structured AI, ship a trader-facing brief.

Exploring how modern data systems separate storage and compute, with a hands-on implementation of a memory engine that works on object stores like S3.

A deep dive into the critical RCE vulnerability affecting Next.js and React 19, with a full reproduction and technical breakdown.