ARMAANΒ AGRAWAL
Software engineer. AI in production. CS @ Northeastern.
Seven stories. Each one real.
π Where it started
IB Diploma 44/45. Tokyo β Boston, alone at 18.
π§βπ« Teaching Assistant
1 year, 4 months teaching CS fundamentals while still a student myself.
π» Code4Community
4 microservices. 5-person team. Led the architecture of a mobile app for Boston nonprofits.
βΎ Boston Red Sox
API latency: 1.2s β 121ms. Journalists at Fenway depended on it.
π United Nations, Geneva
I was 19. The person next to me was negotiating arms treaties.
π’ NExT Consulting
63% faster decisions in chemical plants. When a pump fails, minutes matter.
π AgentOps Hackathon
Built a safe AI agent in 3 hours. Won 2nd place.
Where I Come From
2022 β 2024
Grew up in Tokyo. Moved to Boston alone at 18 to study CS at Northeastern. IB Diploma: 44/45 β top 1% globally.
Teaching Assistant
1 yr 4 moTeaching CS fundamentals while still a student myself.
To explain recursion to someone who's never written a loop β you have to understand it at a different level. Not memorized. Not functional. Understood. Teaching forced that on me for 16 months straight.
Code4Community
My first real team. Built a mobile app for Boston nonprofits. I led the architecture.
microservices. 5-person team. Design to deployment.
The decisions I made would outlive me on the codebase. Every module boundary, every API contract β the next team would inherit them. That was the moment I stopped thinking like a student and started thinking like an engineer.
Beyond the Codebase
Mentored a teammate from zero TypeScript to shipping features independently β pair programming, code reviews, enough context to make decisions alone.
Teaching one person to contribute is a multiplier. You don't just get a task done β you get a teammate who can now do it without you.
The First Time It Mattered
Jan 2024 β Sep 2024
The Boston Red Sox needed live batting lineups for journalists during games. The old method was a handwritten whiteboard. If this went down mid-game, the press couldn't report.
The first time I pushed to production at Fenway, I refreshed the page myself just to make sure it worked. No staging environment that matched. No way to undo. Real users. Real stakes.
Added Redis caching and Celery async workers. Response time: 1.2s β 121ms. Live at Fenway Park.
90% faster means a journalist can report the lineup change before the first pitch. That doesn't sound like engineering. But that's exactly what engineering is for.
Code Has Consequences
Summer 2023
Northeastern selected me for a month at the United Nations in Geneva. I sat with diplomats, Nobel laureates, and officials debating AI regulation and autonomous weapons.
I was 19. The person next to me was negotiating arms treaties. We were both talking about AI. That's not a metaphor. That was the room.
I used to just think about the product.
Geneva changed how I build. Now I ask: who's on the other end? What can they afford to lose? What can't they?
The Hardest Problem
Jul 2025 β Dec 2025
Chemical plants sending thousands of sensor readings per second. The data was messy and inconsistent, in a domain I knew nothing about.
A chemical engineer doesn't care about your tech stack. They care if the pump is going to fail β and whether they'll know in time to stop it.
Faster decisions during critical equipment failures. Operators can act immediately.
Industrial Monitor β Live Sensor Data
63% faster isn't a metric on a slide. It's the difference between catching a fault and missing it.
Students matched with their first career experience.
The Cold-Start Problem
New students have no history. Started with text matching, then shifted to collaborative filtering β βstudents like you also liked this.β Good matches from day one, better over time.
Co-op Recommendation Engine
2,547 matches made β
At Northeastern, co-op is career-defining. A bad match wastes a semester. A good match changes the trajectory of someone's career. The engine found one student a position aligned with her background. She got the role.
What It Means
Mar 2025
Built a working AI shopping agent in 3 hours and won 2nd place β not for speed, but for what I built into it.
3 HRS$ init agent --framework openai-agents-sdk
> Shopper agent scaffolded β
> Adding input guardrails...
> Prompt injection test: BLOCKED β
> Harmful request test: BLOCKED β
> Prototype functional in 3 hours β
> Result: 2nd Place π
Most hackathon demos ship a prototype and call it done. This one had safety built in from line one. Not because the rubric required it. Because that's the right way to build.
Input Guardrails
AI agents can be tricked through prompt injection. I built safety layers that inspect every input before it reaches the AI β blocking attempts to override instructions or extract data.
Anyone can build fast. The question I keep asking is: what happens when someone tries to break it? That question is what separates a demo from a product you can trust.
IΒ SHIPΒ AIΒ TOΒ PRODUCTION.
Real users. Real load. Real stakes. If you're building something that needs to work in production β not just in a demo β let's talk.
Armaan Agrawal Β· CS @ Northeastern Β· Boston, MA Β· ζ₯ζ¬δΊΊ π―π΅ Β· 2026