About · United States
VK
Applied AI Engineer · AI Solutions Architect
I build AI products. Privacy-first agents, kid-safe educational AI, production SaaS with real customers, on-device voice — every piece shipped, every piece used.
I work end-to-end: model architecture and safety guardrails through to payments, OAuth, observability, and the unglamorous middle layer that decides whether a real customer stays. The work on this site spans local LLMs on Apple Silicon, multi-agent systems with row-level ACLs, and a paid commercial engagement with a real client.
What I look for in a role. Senior IC or first-line management on a team that's actually shipping AI to humans — Applied AI Engineering, Solutions Architecture, or AI Engineering leadership. I care about the boring middle (auth, payments, eval harnesses, observability) at least as much as the model layer, because that's where products live or die.
Toolkit
What I reach for
AI / ML
- Anthropic Claude (Sonnet 4.6, Opus 4.7)
- Ollama (qwen3.5, gemma3)
- MLX + speculative decoding
- Groq + multi-model fallback
- pgvector / RAG
- Voice (STT/TTS/VAD, Kokoro, parakeet, Silero)
- Evals & A/B (OpenRouter)
- Prompt caching, system-prompt design
Languages & Runtimes
- TypeScript
- Python
- Node.js 22
- Bun
- Swift / SwiftUI
- SQL
Frameworks
- Next.js 15 / 16
- Astro 5
- React 19
- React Native (Expo)
- FastAPI
- Vercel AI SDK
- Tailwind v4
- Drizzle ORM
Data & Infra
- PostgreSQL + pgvector
- MySQL
- SQLite + sqlite-vec
- Supabase
- Proxmox / LXC
- Docker / Compose
- Cloudflare Tunnel
- systemd / launchd / PM2
Product
- Stripe / Lemon Squeezy
- OAuth (Google, Apple)
- Chrome Extensions (MV3)
- WordPress / Kadence
- SEO + JSON-LD + llms.txt
Philosophy
How I work
Ship the thing. A polished prototype that nobody uses is worse than a rough product that one real user relies on. Every project on this site has a real user, even if that user is my mother.
Treat AI like a colleague, not a magic box. The model layer is one engineer in a much bigger system. Auth, payments, observability, safety guardrails, evals — that's where the real work hides, and where the LLM is at its most useful as a builder.
Privacy and safety aren't features. They're shapes the architecture has to take from the start. Three-tier safety classes, row-level ACLs, on-device inference where it matters — bolted on later, none of these work.
Measure what you change. Vibes are not a roadmap. The reason PostyPop has an audit log is the same reason the Lead Engine tracks every preview click: you can't iterate on what you can't see.
Why Claude
I bet on Anthropic, and the work shows it.
I built this site with Claude Code, and I use Claude every working day across these surfaces:
- Claude Code (CLI) as my pair programmer for every project on this page. Multi-file edits, planning, refactors, and the actual git workflow.
- Claude API in production runtime for Jarvis V5, Lead Engine, Prerna Kapoor's site, and the personalized cold-outreach drafting.
- Claude Skills for project-specific workflows (humanize pass, PostyPop brand voice, Foundation Implants tone).
- MCP servers for tool access from inside the Claude Code session (filesystem, git, GitHub, Telegram).
- Sonnet 4.6 and Opus 4.7 picked deliberately per task: Sonnet for high-volume work, Opus when the architecture or judgment matters more than the latency.
The reason I default to Claude isn't loyalty to a brand. It's the model itself, and it's the alignment work that goes around it. Privacy-first agents, kid-safe tutors, three-tier safety guardrails — the projects on this page are built on assumptions Anthropic's research papers explicitly defend. Picking the model that takes those questions seriously is the same call I'd make as a hiring manager picking a team.
That's why I'm applying.