MS @ NYU · Dec 2026
ARYAN THEPADE
I ship AI products that solve
real business problems.
Built $4M agritech startup from zero.
2000 farmers. 4 supply chains across
4 states in 2 months.
Built $4M agritech startup. Coordinated farmers, buyers, logistics across India. Deployed ML in production.
Frontier AI @ NYU. Transformers from scratch. RLHF. Agents. Production ML.
Build AI systems that ship to production. Not demos—products people use.
THE FASALTECH STORY
Summer 2020
Game of Thrones was ending. Outside my window, farmers were protesting.
I felt guilty. Had zero knowledge about farming. Started anyway.
Everything failed.
App: 12 downloads. Pesticides: 0 sales. Wasted months.
February 2022
Farmer on TV selling muskmelons for 10x normal prices.
"Premium crops need premium buyers."
Built win-win model:
→ Farmers: premium prices
→ Buyers: quality supply
→ Me: coordination margin
First deal: 5 tons. It worked.
September 2021
Lulu Dubai: "We need 50 tons of watermelons."
Maharashtra was flooding. Zero contacts in Karnataka.
Went anyway. Built 4 supply chains across 4 states in 2 months.
Karnataka · Andhra Pradesh · Telangana · Tamil Nadu
Delivered on time. They kept ordering.
India's first seedless watermelons at commercial scale
Clients: Lulu, Reliance, Namdari, Zepto, Swiggy
WHAT I'M LEARNING NOW
$ cat current_focus.txt
BUILDING LLM REASONERS @ NYU
with Greg Durrett
CURRENT TOPICS:
→ Transformers architecture
How attention actually works
Status: ████████████░░ In progress
→ FlashAttention
Making transformers fast
Status: ██████░░░░░░░░ Learning
→ RLHF & Alignment
Teaching models to follow
Status: ████████████████ Deep dive
→ Chain-of-thought reasoning
Making models think step-by-step
Status: ████████░░░░░░ Exploring
→ Agentic systems
Models that plan & execute
Status: ██████░░░░░░░░ Starting
FROM FIRST PRINCIPLES.
NOT JUST API CALLS.
[2026-01-26] Current:
FlashAttention implementation ✓
[View weekly updates →]
■
COURSEWORK
Building LLM Reasoners
Greg Durrett, UT Austin
Topics: Transformers, attention mechanisms, RLHF, reasoning, agents
Leadership & Stakeholder Management
Organizational behavior, team dynamics, decision-making
Lean Launch Pad
Customer development, market sizing, business validation
SELECTED PROJECTS
Pico LLM Interpretability
Understanding how small language models form representations.
Mechanistic interpretability on 10M parameter models. Discovered functional roles: early layers handle syntax, late layers handle semantics.
Similar to Anthropic's interpretability work, but on models small enough to fully understand.
FasalTech ML Production System
Disease prediction for 2000 farmers with terrible connectivity.
Not model accuracy—it was 2G networks, low-quality cameras, farmers who'd never used AI.
72% accuracy vs 65% for expert agronomists. Available 24/7 vs 10 farm visits/day.