©2026 / New York, NY

Hi, I am

Bryan
Mejia

Portrait of Bryan Mejia

I am a software engineer in New York. Today I’m a Cloud FinOps Engineer at the American Arbitration Association, governing spend across eight cloud and AI vendors. Before that I led a dev team for a contextual memory and learning platform at an agentic AI startup, orchestrating multi-agentic systems for agnostic agent harnesses. My focus is AI and agentic systems: designing systems that adapt and learn in dynamic environments, and solving complex business problems.

See the Work Resume
About

From computer science principles to agentic systems in production.

My path through software engineering has been additive. Each layer I learned became the foundation for the next, and all of it now points at AI.

It started with computer science principles at CUNY Queens College: data structures, algorithms, operating systems, and the habit of reasoning about a problem before reaching for a keyboard. That grounding is still the part I lean on most.

From there I moved into software engineering proper, learning how systems are actually built and maintained rather than merely designed. Java and C++ taught me structure, JavaScript and TypeScript taught me the web, and shipping to real users taught me the difference between working and reliable.

Data science and machine learning came next. Working with Pandas, NumPy, and model pipelines shifted how I think about problems, from deterministic logic toward evidence, evaluation, and uncertainty.

That path led directly into AI and agentic systems, where I work now: agent harness integration, lifecycle-hook context injection, tool calling, RAG, and the cloud infrastructure and cost engineering that keep it all running in production.

Leadership

  • Fellow

    Google Software Engineering Fellowship, BASTA · Mar › May 2026

    Competitively selected from a national applicant pool for a Google-backed software engineering fellowship, with direct mentorship from a Google Cloud Engineer.

  • Treasurer and Event Coordinator

    Society of Industrial and Applied Mathematics, Queens College · Feb 2024 › Jan 2025

    Allocated over $7,500 in combined budgets for professional development initiatives and hackathon sponsorships serving 200+ members.

    Organized Python and math workshops, driving a 70% increase in active member participation over two semesters.

Stack

What I work with

Languages and Frameworks

  • Python
  • TypeScript
  • JavaScript
  • Node.js
  • Express.js
  • React.js
  • Java
  • Spring Boot
  • C++
  • FastAPI
  • Pandas
  • NumPy

Cloud and Databases

  • Microsoft Azure
  • Power BI
  • Oracle Cloud
  • Google Cloud
  • Kubernetes
  • Helm
  • Docker
  • OpenTelemetry
  • Datadog
  • Cloudflare
  • PostgreSQL
  • Firebase
  • Git
  • GitHub

AI Tools

  • Anthropic Claude
  • Google Gemini
  • OpenAI Codex
  • Claude Agent SDK
  • MCP
  • Agentic workflows
  • Harness integration
  • Lifecycle hooks
  • Tool calling
  • RAG
  • Evaluation frameworks
  • Prompt engineering
Professional Experience

Where I have worked, and what I built there

Jul 2026 › now

Cloud FinOps Engineer

American Arbitration Association, New York

  • Own cost governance across eight cloud and AI vendors, among them Oracle Cloud, Azure, OpenAI, Anthropic, ChatGPT, and Datadog, engineering a system to cut AI token consumption and maximize prepaid credit utilization.
  • Built an executive Power BI dashboard that unifies all eight vendors on a single cost model, landing Oracle and Azure as FOCUS exports and pulling the remaining six through their own APIs to normalize against them. Refreshed twice daily, it shows services, usage, variance, run-rate forecast, and trends over the full inventoried history in one view that leadership previously assembled by hand across eight separate pricing models.
  • Designing an agentic optimization system across 15 OpenAI projects and a ChatGPT workspace, routing to lower-cost models where quality permits and pacing usage against prepaid commitments to avoid both credit forfeiture and overage.
  • Built the cost data foundation on the FOCUS open billing specification, cataloging 40+ billing scopes across OCI compartments, Azure subscriptions, and OpenAI projects with ownership, cost model, and commitment terms.
  • Oracle Cloud
  • Azure
  • OpenAI
  • Anthropic
  • Datadog
  • Power BI
  • FOCUS
$170K
monthly spend governed
8B
monthly tokens
100+
hrs saved monthly

Jan › Jul 2026

Forward Deployed Engineer

Augenta AI, New York

  • Led the engineering team and assisted with infrastructure for an Azure cloud-native contextual memory platform deployed in production for enterprise customers.
  • Developed a plugin for three agent harnesses, Claude Code, Codex, and GitHub Copilot, capturing agent trajectories and user-agent interactions across all coding sessions, streaming telemetry through an OpenTelemetry to Event Hubs producer/consumer buffer into a downstream memory engine.
  • Engineered lifecycle-hook context injection that reused shared context across users, agents, and sessions.
  • Azure Microservices
  • EventHubs
  • Claude SDK
  • Codex SDK
  • GitHub Copilot
  • Data Pipelines
  • OpenTelemetry
24%
token usage cut
33%
agent recall gain
3
agent harnesses supported

Jun 2022 › Apr 2026

Operations Assistant

MQ Steel Corp, New York

  • Optimized cross-functional project workflows across 25+ active client accounts, reducing average turnaround time through process standardization and task coordination.
  • Maintained a client database for 25+ accounts with 100% accuracy in financial record-keeping.
  • Operations
  • Process design
  • Data accuracy
15%
faster turnaround
$500K+
annual transactions
25+
client accounts
Selected work

Things I have shipped

  • Abstract diagram: markdown notes feeding a derived index that returns a ranked answer, top result highlighted
    01 / Retrieval and memory system Aug 2026

    Self-Learning Agentic Brain

    A knowledge system an agent can actually retrieve against. Markdown is the canonical store and every index is derived and disposable, so the whole thing rebuilds from source and returns identical results. Lexical BM25 retrieval sits behind a relevance gate, and a golden-set harness scores every change on precision@1, recall@k, and MRR, so enrichment is judged by whether the number moves rather than by feel. A neutral control set validates the benchmark itself: it caught the tracked set sharing vocabulary with its own note aliases and inflating results, which voided an earlier verdict against dense retrieval. The engine is open source; the vault it indexes stays private.

    65.7%
    precision@1
    88.6%
    recall@k
    35
    golden eval cases

    Python / BM25 / SQLite / Eval Harness / Knowledge API

    View Repository
  • Abstract diagram: a spoken loop, voice into a terminal session hosting an agent and the reply spoken back
    02 / Automated Technical Operations Manager Aug 2026

    ATOM

    A voice-enabled terminal suite that hosts an AI coding agent in a chromeless desktop window, closing a full spoken loop: tap to talk into the session, and the finished reply is read back aloud in a chosen voice. Input rides the agent’s own tap mode; the spoken output is a stop hook I wrote that strips code blocks, caps the utterance, cancels the previous one mid-sentence, and switches voice by language, because an English voice reading Spanish is unintelligible rather than accented. Underneath, a real PTY is bridged over a WebSocket into xterm.js while four live endpoints drive the widgets around it. Sessions run under the same hook layer as my knowledge vault, so each prompt is answered with retrieved context and captured back into it afterwards.

    4
    live widget endpoints
    2
    spoken languages
    0
    bundled Chromium

    Node / node-pty / WebSocket / xterm.js / Python / tmux

    Private Repository

  • Abstract diagram: an array being stepped through, with a tree above it and a scrubbable timeline below
    03 / Interactive DSA platform Aug 2026

    Structa

    A gamified way to learn data structures and algorithms: every structure is a machine you can poke, every algorithm a scrubbable step animation with a narrator. Interactive DSA topics are each taught end to end in Python, C++, and Java, alongside SWE, system-design, API, and database guides, a curated LeetCode ladder, and game worlds with three difficulty tiers. Progress, XP, and achievements persist in the browser.

    19
    interactive DSA topics
    295+
    curated LeetCode problems
    11
    game worlds

    TypeScript / Cloudflare Pages / Canvas

    View Live Site
  • Abstract diagram: a public intake form feeding a locked inner portal where documents are drafted and sealed
    04 / Business Operations RAG Platform Jun 2026

    MQ Steel Platform

    A request-intake site and bilingual admin portal built for an industrial welding company that was still taking jobs by phone. Submissions are validated in Firestore security rules rather than client code, so the browser is never trusted, and staff triage them in English or Spanish with automated email on every status change. Inside sits Flux, a retrieval-grounded assistant that drafts invoices and faxes from the company's own records — the model writes the prose, integer-cent math and an atomic counter own the money.

    60+
    welding portfolio images
    2
    portal languages
    0
    third-party trackers

    JavaScript / Firebase / Cloudflare Workers

    View Live Site
  • Abstract diagram: many job postings funnelled down to a short ranked shortlist
    05 / Job-search copilot Jul 2026

    JobBear

    A human-in-the-loop copilot for the new-grad software engineering job search. It discovers live openings across public applicant tracking systems, filters them down to genuinely eligible roles, ranks them against a resume with a local model, and remembers what it has already seen so each day surfaces only what changed.

    ~500
    company slugs probed
    100%
    local ranking model

    Python / Local LLM / ATS integrations

    Private Repository

  • Abstract diagram: three agents reading a source file, with three lines flagged as findings
    06 / AI-driven code auditor Jan 2026

    Security Sentinel

    A multi-agent system that statically analyzes large-scale Python repositories, using tool calling and structured LLM outputs to identify logic vulnerabilities and code smells. Agentic RAG injects project-specific context into prompts to reduce hallucinations.

    85%+
    detection accuracy
    30%
    precision over zero-shot

    TypeScript / Anthropic Claude API / Multi-agent RAG / Docker

    View Repository
  • Abstract diagram: a camera feed and an audio feed merging into a single connection into a service where three models run in parallel, scored output returning along an evaluation loop
    07 / Multimodal debate coach Oct 2025

    Polly Debate AI

    A debate and public-speaking coach that watches, listens, and scores you in real time. Camera frames and audio stream over one WebSocket to a Python service running YuNet face detection and FER+ emotion classification as ONNX models, MediaPipe hand tracking, and librosa pitch and energy analysis, with Gemini handling transcription and coaching. The emotion pipeline shipped unmeasured, so I built the evaluation harness that measured it — then replaced both stages, and dropped two classes the numbers showed were unreliable rather than surfacing them anyway.

    175
    labeled faces
    37->51%
    balanced accuracy
    688MB
    dependencies, was 2.1GB

    Python / FastAPI / ONNX / Cloudflare Pages / WebSockets

    View Live Site

Project 1 of 7

Off the map

Places I have been

Drag, scroll, or use the arrow keys

  • A busy Tokyo crossing framed by tall buildings covered in illustrated billboards
    Tokyo Japan Temples and skyscrapers, plus Osaka, Kyoto, and Nara.
  • The Trevi Fountain in Rome, its sculptures and turquoise basin crowded with visitors
    Rome Italy First time in Europe. The Colosseum, the Vatican, the food.
  • A sunlit Sardinian street of pastel buildings with green shutters and Italian flags
    Sardinia Italy The Mediterranean coast, local towns, a lot of seafood.
  • Two red telephone boxes on a paved London street
    London England The Tower of London, Big Ben, the British Museum.
  • A wide canyon overlook with a river bending far below
    Arizona United States Hiked the Grand Canyon, away from the city for a while.
  • A giant pineapple sculpture standing beside a roadside in Ecuador
    Guayaquil Ecuador Visiting family and exploring my roots.
  • A white stone lion monument in a marble hall, inscribed with Civil War battle names
    Boston United States HackHarvard 2024 and 2025.
  • Bryan in cap and gown at his college graduation, holding bouquets beside a fountain
    New York United States Home base. Studied here, built here.

1 of 8

Contact

Get
in
touch

Say Hello