AaradhyaKhanna

AI/ML engineer. I build machines that see, decide and act in the real world.

View works Get in touch

I'm Aaradhya Khanna, an AI/ML engineer from Punjab, studying Computer Science Engineering (AI/ML) at Chitkara University. I like problems where a model has to meet the real world: a road, a vehicle, a supply chain, a company full of agents.

In September 2026 I co-founded MAARG Technologies, which builds AI and hardware systems for infrastructure; I'm Technical Lead at PANEVO, where I built and run the production web platform; and I lead AI for Team Ashvamedha, my university's autonomous vehicle team for aBAJA under SAEINDIA.

Before that I worked on voice cloning and retrieval at ZSapiens and on end-to-end ML pipelines at SkillAiro, and studied applied AI at FH Aachen in Germany. I also write: a research paper on reversal asymmetry in GPT-2, two applied ML studies and an essay on recommendation systems.

Education · Aug 2025 to 2029

B.E. Computer Science Engineering (AI/ML)

Chitkara University, Punjab, India

  • Data Structures
  • Algorithms
  • Machine Learning Fundamentals
  • Artificial Intelligence

CGPA

9.94

out of 10

Experience & achievements →
Computer visionAutonomyAgent organisationsRAGVoice AIEdge hardwareInterpretabilityData engineering

Since Sep 2026

Co-Founder

MAARG Technologies Pvt. Ltd.

Co-founded a company building AI and hardware systems for real-world infrastructure, starting with MAARG Intelligence: a smart traffic platform for signals, cameras, sensors, message boards and operators.

Since Sep 2026

AI Lead

Team Ashvamedha · aBAJA, SAEINDIA

Lead the AI work stream of Team Ashvamedha, my university's team for the aBAJA competition under SAEINDIA: model training and making the vehicle drive itself.

Since Apr 2026

Technical Lead

PANEVO

Lead the company's technical initiatives: architecture, technology stack and engineering decisions.

  1. maarg.company

    MAARG Technologies

    AI and hardware systems for real-world infrastructure, starting with traffic.

  2. panevo.web

    PANEVO

    The production website for a flavoured fresh paneer brand: designed, built and deployed.

  3. abaja.autonomy

    aBAJA 2027

    Making an all-terrain BAJA vehicle drive itself, as AI Lead for Team Ashvamedha.

  4. mycel.network

    Mycel

    Give it a product. An organisation of 22 AI agents designs the supply chain, then keeps watching it.

  5. kibbutz.office

    Kibbutz

    An autonomous AI company in a pixel-art office: 27 agents, 7 departments, all visible while they work.

  6. pinion.studio

    Pinion

    A no-code studio for 3D websites: drag a world together, export real React code.

Reverse accuracy by model size

GPT-2 Small · 117M

9.33% SFT
95.56% ICL k=4

GPT-2 Medium · 345M

64% SFT
98.67% ICL k=4

GPT-2 Large · 774M

40% SFT
100% ICL k=4

Fine-tuning on forward questions reaches 100% forward at every scale, but reverse accuracy collapses even with the fact in the prompt. Four demonstrations bring it back.

Every wrong reverse answer is empty or the forward answer: a learned direction, not a lost fact. Demonstration information reaches the final position between layers 11 and 16 in Medium.

Reversal Asymmetry in GPT-2

Forward-Only Fine-Tuning Induces a Reversal Asymmetry in GPT-2. I fine-tuned GPT-2 Small, Medium and Large on forward-only questions over a synthetic dataset and tested them on held-out facts whose statement is always in the prompt. Every model answers forward questions perfectly but fails many reverse ones, and every wrong reverse answer is either empty or the forward answer: the model has learned a directional answer rule, not lost the association.

Read the case study →Code & data ↗

More experiments

All writing →

Roles and internships, newest first. Clubs, education and certifications are on the profile.
Full profile →

Open a channel.

Research collaborations, robotics and vision problems, internships, or something you are building that has to work outside a notebook.

aaradh.kh@gmail.comWrite a message →