Builder · 20-year-old sophomore @ UMass Amherst

I ship production AI systems — agentic pipelines, applied ML, infrastructure — not demos. I move fast, I teach what I learn, and I'd rather build the thing than talk about building it.

• 1st Place — HF0 AI Hackathon — selected as 1 of 30 builders from 1,000+ applicants; built a winning product in 12 hours with a team I met that morning and demoed live to top VCs and AI engineers in SF.
• Breakthrough Tech AI Fellow — MIT — competitive, fully-funded AI/ML fellowship with real industry projects and mentorship.
• AI Engineering Fellow — ASAPP — shipped an agentic NLP extraction system processing 10K+ multi-turn dialogues at 95–100% accuracy; results presented to ASAPP ML leadership.
• Co-authored AI research — accepted to NUTRITION 2026 (American Society for Nutrition / ASN) — on LLM reliability in clinical dietary guidance.
• Left a Purdue research offer to build full-time — because the worst thing you can do is not show up.

Workflow intelligence for teams: mined behavioral patterns from the tools teams already live in (GitHub, Slack, Linear) and deployed autonomous DAG-based agents to execute discovered workflows. Backed by Founders Inc. Speedrun Top 10%. YC Top 10%.

Agentic NLP system (LangGraph) for structured extraction from unstructured customer conversations. Processed 10K+ multi-turn dialogues at 95–100% accuracy on key fields. Hybrid Regex + LLM-agent pipeline on Gemini 2.5 Flash; benchmarked with PCA/KMeans clustering and KNN classification (F1 > 0.85). Presented to ASAPP ML leadership.

Drag-and-drop neural network builder that makes ML education hands-on. Build, train, and export deep learning models (classification, regression, CNN) with real-time feedback. Used by 500+ high school students.

Evaluated frontier LLMs (GPT-4.5, Gemini 2.5, Claude 3.7, DeepSeek-R1) against WCRF/AICR clinical guidelines for accuracy, consistency, and bias. Built a multi-agent pipeline orchestrating Hugging Face APIs for parallelized insight generation; automated CSV-to-prompt tooling that cut parsing time 60%. (→ research accepted to NUTRITION 2026.)

• Paktolus — Architected a full-stack user management system (Django REST, FastAPI, Next.js 15, TypeScript) with JWT auth and RBAC; 80%+ test coverage.
• telSpiel — Built React dashboards supporting 150M+ daily interactions, cutting navigation time 40% and compliance errors 35%.

• President — CICSoft @ UMass — running the largest tech organization at UMass. My whole job is help students develop skills to land roles at crazy places.

I'm happiest when I'm making something that real people use.

• Competitive chess — 🥇 1st Place, UMass Spring Classic (U1200). I like winning at things, and I like the kind of slow, ruthless thinking it trains.
• I teach. NeuroBlock put hands-on machine learning in front of 500+ high schoolers, and as CICSoft president I spend most of my energy getting other students into rooms they didn't think they could reach.
• Probability & markets — the SIG fellowship rewired how I think about decisions: bet on expected value, size your conviction, and don't confuse a good outcome with a good process.
• I build because I can't not — I'd rather ship a rough v1 today than a perfect plan next month.

"The worst thing you can do is not show up."

Focus: Agentic AI · LLM systems · multi-agent orchestration · on-device inference · production ML.

Let's build something. → aryamansgoenka@gmail.com