What I'm
doing now.
The /now page convention — what I'm actually working on, rather than a résumé.
// 01 — STATUS
Analyst at TQ Ventures — New York, NY. Investment team at an early-stage venture fund. Engineering-flavored role: applied AI, LLM agent systems, and internal tooling. Before this: Founding Engineer at Structured AI (NYC), February to June 2026 — 1,035 commits over a documented 14-week sample, agent infrastructure + ZDR-compliant Python sandbox + iOS app from zero.
// 02 — TALKING ABOUT
Agent infrastructure, MCP, multi-vendor LLM orchestration, AI safety / sandboxing, durable chat, real-time data, and vertical AI for regulated domains.
If you're building any of those — please email tanayshah2024@gmail.com. I respond within 24 hours.
// 03 — BUILDING
Public case studies of agent infrastructure work. Each entry shows the why, the architecture, and the lessons. The most recent additions:
- → Real-Time Sales-Conversation Coaching Agent — two-path architecture (sub-2s live coach + post-call deep dive)
- → Travel MCP Server — flights, hotels, weather as MCP tools with per-tool TTL caching
- → Mercury Stream — real-time market data pipeline with on-anomaly black-box flight recorder
- → AI Agent Error-Handling Patterns — 4 production reliability patterns on Trigger.dev v4
// 04 — WRITING
Field-report blog posts on agent infrastructure. Each is shaped for engineering depth, not vendor marketing.
- 📝 What the Bubblewrap Sandbox Escape Tells Us About Agent Runtime Hardening in 2026
- 📝 Picking MCP Servers for an Agent — A Selection Heuristic for 2026
- 📝 Designing Tool Surfaces for LLM Agents
- 📝 Multi-Vendor Agent Design — Why One Model Isn't Enough in 2026
- 📝 Building a Zero-Data-Retention Layer for Production LLM Agents
// 05 — STUDYING
Million-token context windows and what they actually change about long-context agent design — where they replace retrieval, where they quietly make it worse, and what the lost-in-the-middle failure looks like in production. Agent runtime hardening as a security discipline of its own: the 2026 Bubblewrap sandbox escape and the Semantic Kernel and Claude Code Hooks CVEs are the same lesson from three directions. Production agent eval methodology; Apache AGE for graph memory; how vertical-AI startups structure training pipelines.
// 06 — NYC AI EVENTS AHEAD
- → AI Engineer NY — October 12–14 2026, NYC
// 07 — LOCATION
New York City. Time zone: ET.
// 08 — STACK ON DECK
Python 3.12 (asyncio TaskGroup) · TypeScript 5.5+ · Swift / SwiftUI · Anthropic Claude (Opus / Sonnet) · Google Gemini 3 Pro Vision · Llama 3.3 70B (via Groq) · LangGraph · Anthropic Claude Code SDK · Trigger.dev v4 · Model Context Protocol · FastAPI · PostgreSQL (asyncpg, pgvector) · Redis · gRPC + Protobuf · bubblewrap + seccomp · Apache AGE · Docker multi-stage · Vercel · Azure Container Apps Jobs.