I'm an AI engineer with a decade+ building production systems in financial services — now designing GenAI and agentic systems end to end, from RAG and eval harnesses to the CI that ships them.

AI Engineer / New York, NY

NewAI Systems2026

Agent Memory

Durable cross-repo memory for Claude Code, GitHub Copilot, and Codex. Each client engagement is a separate store rather than a filtered view, so one engagement's notes are unreachable from another even by exact id. Knowledge exports at global scope by default — what you learned travels between machines, the client's architecture does not. Zero runtime and zero dev dependencies, asserted in CI and published to npm with Sigstore provenance attestation.

Projects

14

0

Runtime deps

93

Tests

12

CI jobs

PythonFastAPIRAGAgentic AIEvalsLangGraphLiteLLMQdrantNeo4jOpenSearchAWSKubernetesKafkaMCP

/ About

A decade+ of backend. Now building agents.

I'm an AI engineer with a decade+ building production backend systems in financial services — trading, money movement, and securities management — including four years embedded at Fidelity Investments as an EY consultant, delivering event-driven microservices and high-traffic investment platforms.

Now I design and ship GenAI and agentic systems end to end: tool-calling agents, multi-step orchestration, and RAG with hybrid retrieval across dense, BM25, and graph signals. I build the evaluation harness that gates promotion — golden datasets, versioned baselines, regression gates — so every retrieval and model change is measured before it ships. I also lead AI enablement for 200+ engineers across the institutional vertical.

/ Toolkit

Languages
Python, TypeScript, Go, Java, SQL
GenAI & Agents
RAG, Hybrid retrieval, LangGraph, CrewAI, MCP, LiteLLM, OpenAI SDK
Evaluation
Golden datasets, Regression gates, Guardrails, PII-aware handling
Backend
FastAPI, asyncio, Kafka, SSE streaming, Microservices
Infra & Data
AWS, Kubernetes, Terraform, PostgreSQL, Qdrant, Neo4j, OpenSearch

/ Selected work

What I've shipped

AI Systems2026

npm Package

Agent Memory

Durable cross-repo memory for Claude Code, GitHub Copilot, and Codex. Each client engagement is a separate store rather than a filtered view, so one engagement's notes are unreachable from another even by exact id. Knowledge exports at global scope by default — what you learned travels between machines, the client's architecture does not. Zero runtime and zero dev dependencies, asserted in CI and published to npm with Sigstore provenance attestation.

Node.js · node:sqlite · Agent Skills

0

Runtime deps

93

Tests

12

CI jobs

Sigstore

Provenance

AI Systems2025

RAG Platform

Code Analyzer

A code-intelligence platform that ingests and queries large multi-repository codebases — 600+ repos at its first engagement — to power software-engineering audits. A tri-store backbone pairs Qdrant for dense semantic search, Neo4j for call graphs and symbol references, and OpenSearch for BM25 exact-match on the rare identifier tokens embeddings handle worst, fused with Reciprocal Rank Fusion and cross-encoder reranking.

Python · Qdrant · Neo4j · OpenSearch · LiteLLM

600+

Repositories

73%

Keyword recall

520ms

p95 latency

3

Retrieval stores

AI Systems2026

AI Agent

Costco Price Match

An AI agent that scans Costco receipts, cross-references each line item against active deals and temporary price drops, and emails a weekly price-adjustment report. Receipt parsing runs on Amazon Nova; the weekly run is an AgentCore Runtime triggered by EventBridge. Ships with a native SwiftUI companion app — CostScanner — on the App Store.

AWS Lambda · Amazon Nova · AgentCore · SwiftUI

Nova

Receipt parsing

Weekly

Run cadence

AgentCore

Runtime

App Store

Companion

Apps2026

iOS App

Yantra: Second Brain

A native iOS app for capturing voice memos, notes, journals, meeting recordings, photos, and saved URLs — then chatting with them. Transcription, embeddings, and summarization all run on-device via Apple Speech, NLEmbedding, and iOS 26 Foundation Models; only the retrieved snippets and the question leave, sent under the user's own Anthropic key through a stateless proxy that persists nothing. Captures English, Hindi, and Hinglish with translation.

SwiftUI · Apple Speech · Foundation Models · Anthropic

3

Languages

100%

On-device NLP

iOS 26

Foundation Models

0

Bytes retained

/ Experience

Where I've spent my time

Building the retrieval platform behind engineering audits, and the agents that automate enterprise onboarding.

  • Built Code-Analyzer, a code-intelligence platform that makes large commercial financial-services codebases auditable on a consulting timeline, treating retrieval over code as a problem distinct from retrieval over prose.
  • Built the evaluation harness that gates promotion — 30 hand-curated Q&A pairs across 15 indexed repositories, scored on keyword recall, file-pattern hit rate, cross-module recall, and citation rate against a versioned baseline. Hybrid retrieval holds 73% keyword recall at 520 ms, with 100% of answers citing sources.
  • Designed the TerraCore Onboarding Agent — an agentic system driving Terraform, Jenkins, and account-servicing APIs as LLM tools, cutting application onboarding from weeks to hours with resumable, auditable workflow state.
  • Shipped the agent as an async FastAPI service with SSE progress streaming, and made destructive infrastructure operations safe to automate through input validation, allow-listed actions, dry-run modes, and human-in-the-loop approval gates.
  • Leads AI enablement for 200+ engineers across the institutional vertical, covering agentic frameworks and org-scale GitHub Copilot adoption.

/ Education

2019 — 2020M.S., Computer ScienceUniversity of Texas at San Antonio
2009 — 2013B.E., Computer Science and EngineeringVisvesvaraya Technological University

/ Certifications

  • Databricks Accredited Generative AI Fundamentals
  • Developing Serverless Solutions on AWS

/ Writing

Notes from building

Posts

31

Let's build

something good.

Have a project, a role, or an idea worth chasing? My inbox is open — I read everything and reply to most.