/uses

The tools, hardware, and services I reach for day-to-day. Inspired by usesthis.com and the wider /uses tradition. Updated when something meaningfully changes — not on a schedule.

Editor + IDE

Primary

  • VS Code + Claude Code CLIEveryday driver. Claude Code (Opus 4.7 1M) handles the heavy lift; VS Code is the surface I read + edit in.
  • CursorFor Python + notebook work where inline completion is worth the extra latency.
  • JetBrains PyCharmWhen I need real debugging in bigger PyTorch projects.

Fonts + theme

  • JetBrains MonoEditor + terminal + code blocks on this site.
  • Tokyo Night (Storm)Dark theme; keeps contrast crisp under bright lab lights at CMU.

Terminal + shell

Shell

  • Git Bash on Windows (Zsh on Mac)The dev laptop is Windows; my home rig is macOS.
  • Windows TerminalTabs + acrylic + JetBrains Mono.
  • starshipPrompt — git status, kubernetes context, python venv, python version.
  • zoxidecd replacement — matters more than any other single dotfile.
  • fzf + ripgrep + fdFuzzy history / file search / grep replacement. Muscle memory now.

Docker + Kubernetes

  • Docker Desktop + ColimaColima on Mac when I want the VM to not eat 8GB of RAM at idle.
  • k9sCluster TUI; makes reading logs / describing pods bearable.
  • lazygitInteractive git — cherry-picks, rebases, stashes without the mnemonic tax.

AI stack (models + hosting)

Everything vends through CipherStack — my self-hosted key vault at cipherstack.kaushik.cv — with LRU rotation and 60s cooldown on rate limits.

LLMs I actually use

  • Claude Opus 4.7 / 4.8Coding + long-context reasoning. 1M context tier for repo-wide refactors.
  • Groq — Llama 3.3 70B versatileVoice-mail agent LLM (Dyx). LPU sub-200ms first-token latency.
  • Gemini 2.xCheap high-volume classification; 8 keys in CipherStack for headroom.
  • GPT-5Sanity check when Claude and I disagree on architecture.

Voice pipeline (Dyx voicemail)

  • LiveKit Cloud AgentsWebRTC + agent orchestration. voicemail.kaushik.cv is one LiveKit room + one long-lived Python agent.
  • Deepgram nova-3Streaming STT; 300ms endpoint silence, ~180ms transcript latency.
  • Cartesia sonic-3Streaming TTS; first-audio in ~150ms.

Retrieval + vectors

  • QdrantProd vector store since the SAP RAG work. HNSW through 2M+ docs (see the /blog HNSW-vs-IVF-PQ post); revisit IVF-PQ or a quantized two-tier index at 100M+.
  • pgvectorFor projects small enough that a second DB is overkill.

Hardware

Daily

  • MacBook Pro 14" M3 Pro (18GB)Home + travel; the fanless Colima setup means it's silent through most builds.
  • Lenovo ThinkPad X1 Carbon Gen 11 (i7, 32GB)Windows dev laptop from SAP — kept for the RAM + Linux compatibility.

Peripherals

  • Keychron Q1 Pro (Gateron Brown)Tactile, hot-swappable, wired-first — I don't trust Bluetooth for a keyboard.
  • Logitech MX Master 3SOnly mouse I've owned for four years.
  • LG 32UN650 32" 4KExternal for the ThinkPad; USB-C power + HDR at the CMU dorm.
  • Sony WH-1000XM4Library + lab noise cancellation; the M5 is not enough upgrade to justify.

Services I pay for

Infrastructure

  • Vercel (Pro)kaushik.cv + cipherstack.kaushik.cv + voicemail.kaushik.cv. Preview URL per PR is worth the tier alone.
  • Cloudflare (free)DNS + WAF for kaushik.cv apex + subdomains.
  • GitHub ProPrivate repos, Codespaces on the odd Chromebook day.
  • NamecheapWhere kaushik.cv lives. Boring choice, correct choice.

Productivity

  • 1PasswordFor anything not vended through CipherStack.
  • ObsidianSecond brain — daily notes, class notes, drafts of the essays that end up on /blog.
  • Raycast (Mac) / PowerToys Run (Win)Cmd-Space is the OS to me.

What I'm learning

The list changes every few weeks — snapshot as of 2026-07.

Currently reading / doing

  • CMU 11-667 — LLM SystemsFall '26 course; covers inference optimization, serving, distillation.
  • The Rust Programming LanguageThird pass. This time it might stick.
  • Designing Data-Intensive ApplicationsRe-reading with the vector-DB chapter in mind.

Missing something you'd expect to see here? Ping me: [email protected].