Overview
GitHub Contributions
I’m Mateo, an Electrical & Computer Engineering student at Cornell building the infrastructure that serves large models—and the research that stress-tests them.
01InfrastructureSite ↗02ResearchSite ↗03CybersecuritySite ↗
PrivateMind
OpenAI-compatible batch inference, GPU load management, and production text-to-image on B200/B300s.
Open Weight, Open Risk
A training-free jailbreak that steers a target model’s chain-of-thought to expose shallow alignment.
Dispatch
Agentic pentesting platform—and Best Developer Tool at the 2026 Cornell AI Hackathon.
04Edge AI
Groundwork
An always-on maintenance agent that runs entirely on one GB10 and proves its own air gap.
Blog(3)
How marko finds you inside a building
- Published on
- closed-weight ECIopen-weight ECIBCSI (this paper)forecast
Capability curves use Epoch AI ECI for major closed- and open-weight releases, plotted at official release dates. The BCSI series is this post's Bio/Chem Safety Index (BCSI = 1 − mean(BRI, CRI)) under the DeepSeek judge; the hatched region marks BCSI below 0.5.
Stack
Infrastructure
Metrics
Experience
Options Technology
- Location
- New York, New York
- Location type
- (On-site)
Software Engineering Intern, PrivateMind Platform
- Employment Type
- Internship
- Employment Period
- 06.2026—08.2026
- Duration
- 3m
Software Engineering Intern, PrivateMind Platform
- Owned and shipped an OpenAI-compatible REST Batch API end-to-end on cloud-native Kubernetes infrastructure: async GPU workers, exactly-once billing for consistency, DB schema, SDK, Helm deployment, ~7,000 lines of tests.
- Benchmarked frontier LLMs before production serving: Poisson-distributed concurrency sweeps measuring time-to-first-token, inter-token latency, and throughput at p50/p95/p99 across multi-GPU HGX deployments.
- Owned LLM inference gateway load management on NVIDIA B200/B300 GPUs: priority scheduling so batch cannot starve interactive chat, per-model streaming timeouts, and token-budget admission control on GPU memory pressure.
- Deployed a novel text-to-image model to production (GPU config, gateway routing, bounded concurrency, pricing, UI) and built the content-safety guardrail layer: a moderation API and an actionable error taxonomy for opaque refusals.
- Merged 133 PRs across 10 repos in Rust, TypeScript, Python, SQL, and Helm through GitHub Actions CI/CD, building distributed systems for scalability; unified two agent execution paths, deleting ~6,000 legacy lines.
- Shipped a nightly memory-consolidation engine, an automation builder driven by a backend tool-calling loop, chat streaming hardened against state bugs, and an inline co-editing canvas with conflict auto-resolution.
- Linux
- Kubernetes
- OpenShift
- Helm
- vLLM
- SGLang
- GitHub Actions
- Rust
- TypeScript
- Python
- PostgreSQL
Cornell University
- Location
- Ithaca, New York
- Location type
- (On-site)
Engineering Manager & ML Engineer, Generative AI at Cornell
- Employment Type
- Part-time
- Employment Period
- 10.2025—05.2026
- Duration
- 8m
Engineering Manager & ML Engineer, Generative AI at Cornell
- Led an 8-person team on full-stack development of an ESG risk monitoring platform for Investcorp ($60B+ AUM).
- Built a vendor due-diligence agentic data pipeline with fintech company QuickFi, reconciling vendor data from filings.
- TypeScript
- Python
- Agentic Pipelines
- Engineering Management
AI Alignment Undergraduate Researcher, Cornell NLP Group
- Employment Type
- Part-time
- Employment Period
- 10.2025—05.2026
- Duration
- 8m
AI Alignment Undergraduate Researcher, Cornell NLP Group
- Built the AI security evaluation pipeline behind "Open Weight, Open Risk" (co-author): 15 frontier open-weight LLMs across 800 WMDP-derived adversarial bio/chem requests, run locally on B200-class GPUs with vLLM and SGLang.
- Co-designed the training-free inception attack: a small uncensored 7B architect model pivots target reasoning, lifting compliance from ~10% to >95% in 5 iterations, exposing shallow alignment in open-weight reasoning models.
- Python
- vLLM
- SGLang
- LLM Safety
- Evaluation
Hospital for Special Surgery
- Location
- New York, New York
- Location type
- (On-site)
Machine Learning Research Intern
- Employment Type
- Internship
- Employment Period
- 06.2024—08.2024
- Duration
- 3m
Machine Learning Research Intern
- Analyzed immunological datasets with Dr. Amit Lakhanpal (Pandas, NumPy, scikit-learn, PyTorch, TensorFlow, VAEs, t-SNE) to extract rheumatoid arthritis signatures; fine-tuned neural networks, generated ESM embeddings.
- Python
- PyTorch
- TensorFlow
- scikit-learn
- ESM
Education
Cornell University, College of Engineering
- Employment period
- 08.2024—12.2028
- Degree
- B.S.
- Field of study
- Electrical & Computer Engineering
Cornell University, College of Engineering
- Bachelor of Science in Electrical & Computer Engineering, GPA 3.75 / 4.00. Expected graduation: December 2028.
- Coursework: CS 4414 (Systems Programming), ECE 3140 (Embedded Systems), ECE 2300 (Digital Logic and Computer Organization), MATH 2940, PHYS 2213, MATH 2930.
- C
- C++
- CUDA C++
- Python
- Rust
- Computer Architecture
- Machine Learning
Projects(6)
Marko
- Period
- 10.2026
Indoor AR navigation app piloted on a multi-story Cornell academic building. / 2nd Place — MLH Big Red Hacks 2026 · Write-up
- Mapped a multi-story Cornell academic building for routing (Swift, ARKit)
- Solved indoor positioning, where GPS is unreliable, with quick, precise localization from what the camera sees and no beacons or routers, via PyTorch/RoMa and Open3D on RunPod GPUs; added an ElevenLabs voice agent
- Swift
- ARKit
- PyTorch
- RoMa
- Open3D
- RunPod
- ElevenLabs
Groundwork
- Period
- 09.2026
On-premises AI maintenance agent that turns machinery fault signals into cited, verified work orders. / 2nd Place — Dell × NVIDIA AI Hackathon at Cornell 2026 · Write-up · Source
- Co-built the agent loop that turns fault signals into cited, verified work orders
- Served Qwen3.6-35B-A3B via llama.cpp on Grace Blackwell inside OpenShell; its egress policy denied exfiltration
- Qwen3.6-35B-A3B
- llama.cpp
- OpenShell
- NemoClaw
- MCP
- SQLite
- Node.js
- Python
Dispatch
- Period
- 03.2026—Present
Agentic penetration testing platform that turns vulnerability findings into ready-to-merge GitHub PRs. / Best Developer Tool — Cornell AI Hackathon 2026
- Built the backend of a multi-agent security platform: an orchestrator agent reads the target codebase (routes, auth logic, raw SQL), plans targeted attacks, and dispatches pentester workers into isolated Blaxel sandboxes
- Closed the loop from finding to fix: content-hash IDs deduplicate findings, constructor workers patch code and open GitHub PRs, and an Express REST API plus Slack/Linear/GitHub integrations trigger fixes and ticket findings
- TypeScript
- Mastra
- OpenRouter
- Blaxel Sandboxes
- Datadog
- Slack API
- Agentic Systems
- Security
Open Weight, Open Risk
- Period
- 10.2025—Present
Training-free jailbreak in which an uncensored model steers the target's chain-of-thought. / Publication
- Achieves near-universal compliance across open-weight models (DeepSeek, Kimi, GLM, Qwen)
- Iterative prompt injection and LLM-as-judge pipelines to streamline evaluation
- Evaluated 15 frontier models across 800 WMDP-derived biosecurity and chemical security requests
- Python
- vLLM
- SGLang
- LLM Safety
- Red Teaming
- Evaluation
- Research
Awards(5)
Cornell AI Hackathon 2026
- Prize
- Best Developer Tool
- Awarded in
- Received in Grade
- University
- Built Dispatch, an agentic penetration testing platform automating vulnerability detection and remediation (Pitch Deck · GitHub)
- Developed integrations with Datadog, Slack, and Linear for seamless integration with developer workflows
Regeneron WESEF 2025
- Prize
- 2nd Place — Computer Science
- Awarded in
- Received in Grade
- High School
Regeneron WESEF 2024
- Prize
- 2nd Place — Computer Science
- Awarded in
- Received in Grade
- High School
Bookmarks(2)
The Techno-Optimist Manifesto
- Author
- Marc Andreessen
- Category
- Article
- Bookmarked on
The Bitter Lesson
- Author
- Rich Sutton
- Category
- Article
- Bookmarked on