Skip to content

Seeking work that excites me

UZAIR
BEG.

Systems & Applied AI Engineer — SF

I build the stuff that has to be fast and can't fall over. Payments at JPMorgan, edge inference at an NSF lab, agent infrastructure at MultAI, and lately the Linux kernel.

Let's chat ↗EmailGitHub
uzair@portfolio — zsh
50+
engineers on Devvy daily
internal, JPMC
~93%
less migration effort
Chase payment teams
10+
clinics running Paws AI
founding engineer
01

/experience

Each one names the call, not the task.

  1. Software Engineering Intern

    JPMJPMorgan Chase & Co.

    Jun – Aug 2026 · Columbus, OH

    Shipped the assigned build. Also won the internal hackathon and kept going.

    • Built Loan Navigator end to end — a React and Spring Boot platform with SSO and automated CI/CD on AWS that tells branch relationship managers which loan product a small business qualifies for, shipped to a Midwest pilot.
    • Drove development across a 2–3 engineer project and ran live feedback sessions with product managers, managing directors, and the relationship managers using it.
    • Built Devvy, an autonomous multi-agent coding assistant that won the firm's internal hackathon, then reached 50+ engineers in daily internal use and cut development time per task by a reported ~20%.
    • Owned Devvy's Model Context Protocol integration layer across Jira, Bitbucket, and Outlook, a three-tier cross-session memory system with keyword-scoped retrieval, and a checkpointed goal/acceptance-criteria state machine sustaining single runs past 10+ hours across model rate limits.

    the call

    Shipped through a mid-build platform migration

    Deployment was already stood up internally when the target moved to AWS. Two of ten weeks went to access and scope churn. It still shipped.

    ReactSpring BootAWSTypeScriptNode.jsMCP

  2. Undergraduate Researcher

    OSUICICLE — NSF AI Institute

    Jan – May 2026 · Ohio State

    HARVEST, the lab's digital agriculture vision framework — real-time weed detection for ground vehicles.

    • Optimized the HARVEST weed-detection inference pipeline with TensorRT and reduced-precision quantization, targeting real-time throughput on an NVIDIA Jetson Orin Nano Super inside a 25 W power envelope.
    • Benchmarked GPU and multi-threaded C/C++ workloads across the compute continuum from A100 and V100 cluster hardware down to the edge device, profiling throughput, tail latency, and synchronization overhead.

    the call

    The budget was derived, not met

    15–20 mph is the tractor speed I engineered against. The field deployment was still ahead of us when I left. My work was hitting the budget on-device.

    C++CUDATensorRTPyTorchONNXJetson

  3. Software Engineering Intern

    JPMJPMorgan Chase & Co.

    Jun – Aug 2025 · Columbus, OH

    Core Payments Infrastructure. Every Chase payment routes through it — Zelle, wires, all of it.

    • Architected a distributed middleware abstraction decoupling legacy Core Payment Systems from the new Payment Instruction Manager, shipped as a shared Java library that teams link directly into their codebase — handling validation, request transformation, and PIM compatibility in-process across dozens of dependent services, cutting cross-team migration effort ~93%.
    • Prototyped and benchmarked the alternatives first: a Spring Boot connector microservice on AWS EC2, then gRPC over HTTP/2 with Protobuf serialization, measured against a mock Create Transaction path and rejected on serialization and network overhead.
    • Held sub-100ms latency and 99.9% availability targets across Zelle and wire transfer flows, with invalid requests failing fast before ever reaching the core system.
    • Became the standard integration path for payment teams onboarding to PIM, and removed per-team infrastructure cost entirely.

    the call

    Deleted the service I proposed

    I pitched a connector microservice and prototyped it on EC2. Benchmarked it: HTTP serialization was pure overhead. gRPC was faster, still not free. The layer only did stateless validation — so a network hop bought nothing and added a failure domain to a path that had to hold 99.9%. I made it a library instead.

    JavaSpring BootgRPCProtobufAWS

  4. Founding Engineer

    PAWPaws AI

    Aug 2024 – May 2025 · Remote

    Vet documentation platform. Four-person startup, zero to production. I owned the backend.

    • Architected multi-tenant FastAPI services on Firestore with RBAC and per-clinic data isolation for clinical records, taking the backend zero to production as founding engineer at a four-person startup.
    • Built a two-tier caching layer combining an in-process LRU cache with Redis in front of vector-embedding and patient-record lookups, cutting database reads by ~65%.
    • Shipped an asynchronous WebSocket layer on FastAPI with connection pooling and Firebase authentication, sustaining 300+ concurrent sessions under 200ms.
    • Engineered the documentation pipeline around two hard API ceilings — Whisper transcription chunked past the 25 MB upload cap, GPT-4o SOAP notes with map-reduce fallback beyond the context window, and FAISS-backed retrieval chat.
    • Onboarded 10+ veterinary clinics; vets reported saving ~2 hours a day on charting and several took on ~10% more patients per month.

    the call

    Both hard limits were API limits

    Whisper caps uploads at 25 MB and real appointments blow past it. GPT-4o's context couldn't hold a long oncology case. Chunking fixed the first, map-reduce the second.

    PythonFastAPIFirestoreRedisWebSocketsWhisperGPT-4oFAISS

  5. Software Engineering Extern

    SNPSnap Inc.

    Sep – Dec 2024 · Remote

    Remote externship. Pick a concept, build a Lens, defend it to Snap engineers.

    • Built an interactive AR Lens in Lens Studio with real-time input handling, animation logic, and engagement mechanics.
    • Presented product reasoning and a live demo to Snap engineers, defending a concept chosen from observed user interest rather than personal taste.
    Lens StudioJavaScriptAR

  6. Software Engineering Intern

    DRBDRB Systems (Vontier)

    May – Aug 2024 · Akron, OH

    Acquired by Vontier in 2021 — same company, two names. Small team with no DevOps practice at all before I got there.

    • Built the team's first CI/CD pipelines from scratch in Azure DevOps YAML covering every C#/.NET Web API and Angular application, automating builds, tests, and AWS ECR/EC2 deployments.
    • Containerized legacy production services with Docker and standardized environment configuration to eliminate cross-environment drift.
    • Replaced a fully manual release process, cutting release time by ~70% and removing 2+ hours of manual QA per deployment.

    the call

    The variance was the problem

    Manual QA ran two hours to a full day depending on config drift. Long is annoying. Unpredictable is unplannable. Containers made it a known quantity.

    C#/.NETAngularAzure DevOpsDockerAWS ECR
02

/projects

Built outside an assignment.

MultAI

Project · 2026

Testing with startups

Built to work out what agent-to-agent coordination looks like when privacy is the binding constraint rather than an afterthought. Everyone in a company gets one agent holding their context, and agents answer each other's questions directly — so you get the answer without interrupting the person. A few startups are testing it.

01

No orchestrator. That's privacy, not scaling — an orchestrator good enough to route well would need read access to everyone's context, which is the exact boundary the product exists to protect.

02

A responding agent runs tool-less: one turn, persona plus question, no tools. So agent-to-agent is depth-1 by construction, and an inbound request can't make your agent read your private data. Blast radius bounded by architecture, not by prompt.

03

Approvals store the action, not a re-derivation. The gated call is persisted verbatim and replayed on approval. The model is never re-prompted, so what runs is what was approved.

04

The permission boundary is an attention boundary. The only gated tools are the two that cost a human's attention. Reads are free.

HonoCloudflare WorkersSupabasePostgres RLSReact 19Anthropic API

Startups

2

AlphaSevn

Co-Founder · 2025

Shut down

Algorithmic prop-trading system — research, execution, and risk in one pipeline. Backtested beautifully. That turned out to be the problem lol

PythonC++Market Data

Fee Dodger

Co-Founder · 2023

Shut down

Inventory platform for resellers — cost basis, marketplace fees, and what you actually cleared. Killed it when the tooling caught up and the whole product became a day of vibe-coding. No moat left worth defending.

TypeScriptNext.jsPostgres

Systems & Open Source

4

QuantLib ↗

Open Source Contributor · Nov 2025

Merged in v1.41

Shipped a const-notional cross-currency swap bootstrap helper in C++ through four rounds of maintainer review. Fixed foreign-leg collateral discounting and replaced weak assertions with par round-trip tests.

C++SWIGQuant Finance

Linux io_uring ↗

Kernel Contributor · Aug 2026 – Present

Patch in progress

Chasing a fixed-file installation regression in io_uring's MSG_RING SEND_FD path, alongside the reporter who has the bare-metal rig. Reading the install path turned up why: the allocation cache is only ever populated on free, and a first fill into a sparse table frees nothing — so every install in a first fill is a guaranteed miss by construction. Preparing a two-patch series: a dedicated kmem_cache for resource nodes, then bulk refill on miss.

CLinux Kernelperfftrace

AETHER

Personal Project · 2025

Source not yet public

Real-time C++ drone fleet backend. Packed-binary UDP telemetry into seqlock-synchronized fleet state with acquire/release ordering across dozens of drones, streamed to Palantir Foundry.

C++UDPSeqlockPalantir Foundry

OpportunityEdu

JPMorgan Code for Good · Nov 2024

1st Place

Teachers generate AI lesson plans and push them straight to a teaching calendar. Built for Opportunity International. I owned backend, AI, and database.

Next.jsFirebaseOpenAITypeScript
03

/about

I'm finishing CS at Ohio State in December. Most of what I've built sits under the product — a payments path that couldn't take another failure domain, an inference pipeline with 8 GB to work with, an agent runtime that has to check its own work.

I like problems where the honest answer is a tradeoff. The best thing I shipped at JPMorgan was a service I talked myself out of building.

Outside that: lifting, soccer, and unreasonable hours of poker ♠ (net negative, thanks for asking).

Currently

  • ›Building MultAI
  • ›Patching io_uring in the Linux kernel
  • ›Finishing my B.S. at Ohio State

Education

  • B.S. Computer Science & Engineering

    Ohio State University

    Expected Dec 2026

$ cat ~/.stack

Languages

  • Python
  • C++
  • Java
  • TypeScript
  • JavaScript
  • C#
  • SQL
  • Bash

Frameworks

  • Spring Boot
  • FastAPI
  • React
  • Next.js
  • Angular
  • .NET
  • Hono
  • Tailwind

Data & Infrastructure

  • AWS
  • Cloudflare
  • Docker
  • Kubernetes
  • Terraform
  • PostgreSQL
  • Supabase
  • Firebase
  • Redis
  • CI/CD
  • gRPC
  • Linux

Systems & HPC

  • CUDA
  • MPI
  • TensorRT
  • CMake
  • ONNX
  • Git

Applied AI

  • Claude
  • OpenAI
  • MCP
  • PyTorch
  • FAISS
  • OpenCV
04

/contact

Seeking work that excites me. The 30-minute link is the fastest way in.

Emailuzairbeg11@gmail.com↗Phone440-412-6300↗CalendlyLet’s chat — 30 min↗LinkedInuzair-beg↗GitHub@uzi101↗
Based inSF