Skip to content

AI ENGINEER · AGENTIC SYSTEMS · STEVENS '25

Autonomous systemsthat shipthemselves.

I design and build production AI pipelines — browser agents, LLM orchestration, and end-to-end automation that runs without a human in the loop.

See my work
PythonTypeScriptNext.jsLangChainAnthropicOpenAIPlaywrightBrowser-UseFastAPIPostgresRedisDockerAWSPythonTypeScriptNext.jsLangChainAnthropicOpenAIPlaywrightBrowser-UseFastAPIPostgresRedisDockerAWS
VercelLangGraphpgvectorPineconetRPCReactTailwindFramer MotionThree.jsn8nZapierGitHub ActionsSESVercelLangGraphpgvectorPineconetRPCReactTailwindFramer MotionThree.jsn8nZapierGitHub ActionsSES

01 — Selected Work

Things I've shipped.

View all on GitHub

Browser-Use Agent for E-commerce QA

Autonomous testing agent that crawls product pages, validates checkout flows, and files Linear tickets for regressions. Cut QA cycle from 3 days to 4 hours.

PlaywrightLangGraphAnthropic

Multi-agent Research Pipeline

Orchestrates 6 specialized LLM agents (search, summarize, fact-check, synthesize, cite, draft) to produce publication-ready research briefs.

LangChainpgvectorNext.js

Voice-to-Workflow Automation

Converts spoken meeting notes into structured Asana tasks with owners, deadlines, and Slack pings — all from a single 30-second recording.

Whispern8nOpenAI

AI Inbox Triage System

Reads, classifies, drafts, and schedules replies for 200+ daily emails. Hits 94% accuracy on category routing in production.

ClaudeGmail APIFastAPI

02 — What I Do

Build the boring parts of AI well.

Production-grade infrastructure for the agentic era — not demos, not slideware.

Agentic systems

Multi-step LLM workflows with memory, tool-use, and recovery. Built to run unsupervised in prod.

Browser automation

Headless agents that navigate, fill, click, and extract — across auth walls, captchas, and dynamic SPAs.

RAG at scale

Vector pipelines that don't fall over at 10M docs. Hybrid search, reranking, and provenance baked in.

LLM ops

Prompt versioning, eval harnesses, latency budgets, and cost guardrails. The plumbing nobody loves.

Edge deployment

Sub-200ms inference at the edge. Vercel, Cloudflare, and self-hosted options.

AUDITPROTOTYPEHARDENSHIP

Workflow design

The 30-min audit that turns "we should automate this" into a shipped system in < 2 weeks.

03 — How I Work

From “what if” to “in production” — in four moves.

01/ 04
01

Diagnose.

A 30-min audit. We map the actual workflow, find where humans rubber-stamp, and identify the single highest-leverage thing to automate.

02

Prototype.

Working agent in 5–7 days. Not a deck. Not a Loom. A real, runnable system you can break.

03

Harden.

Eval suite, error recovery, observability, cost ceiling. The version that doesn't page you at 3am.

04

Ship & extend.

Deploy to your stack, train your team, and design the next three workflows worth automating.

Deep Amish Shah headshot

04 — About

I started writing scripts to skip my own homework. Now I do it for companies.

I'm Deep — a Stevens '25 AI engineer based in Jersey City. I came to AI through automation: every job I've had, I've ended up rebuilding the workflow into something that runs while I sleep.

Now I help teams ship agentic systems that actually hold up in production — not the kind that demo well and break on Tuesday. If you've got a workflow that eats your team's week, I'd like to take a look at it.

  • 12+production agents shipped
  • 200K+LLM calls per day handled
  • 94%avg. agent accuracy in prod
  • <2 wksprototype to deployment
Read the full story

05 — Let's Talk

Let's build something that runs itself.

Move your cursordeep@deepshah.tech