
Founder + AI systems researcher
Shivansh Fulper
I’m building the agent that stays on your side.
Waldo + Kennel · AI systems researcher · B.Tech, IIITDM Jabalpur ’26
Things I’m building.

Models → Agents → World.
I started with the layers below the model, then watched models become agents with tools, memory, permissions, and real users. Now I’m trying to understand what changes when those systems act in the world. Systems Around Models is where I write down what I’m learning about the harness around the model—and what it takes to know the work is actually done.
Explore the fieldbook Opens in a new tabAbout Me
I’m building Waldo and Kennel: a user-owned personal agent and its first home on the Mac. I care about what happens after an agent says “done” — whether the outcome is real, what still needs human judgment, and how one agent can remain on the person’s side across models, tools, work, and life.
At Atlan, I helped build and operate more than 30 production agent instances through AtlanClaw. That work made one gap impossible to ignore: an agent finishing a task and the task actually being done are two different things. It shaped how I think about continuity, permission, evidence, and personal agency.
I started coding at 12 to build a Pokédex, then spent years jailbreaking phones and tracing systems past their intended limits. At IIITDM Jabalpur, I grew HackByte from an internal college event to 5,154 registrations as a lead organiser across three years. Project EKA, Code for GovTech, and rebuilding Qwen3 MoE took that instinct into models and real-world systems. MIRAI-Setu took me across Japan in 2025 and deepened my interest in craft, manufacturing, infrastructure, and long time horizons.

Notes from the work.
Research questions usually arrive after something breaks, surprises me, or refuses to fit the model I had in my head.
- 01research note6 min
When an Agent Says Done, What Is Actually True?
A practical way to separate an agent’s activity from an outcome a person can safely accept.
- 02research note6 min
Memory Is Not Storage
Why a personal agent needs provenance, correction, and forgetting—not an endless archive.
- 03research note7 min
The Harness Is Part of the Agent
What studying more than 40 agent harnesses taught me about the systems around the model.
What each chapter made me ask.
Founder
WaldoBuilding a user-owned personal agent and its first Mac home, while keeping product claims separate from the deeper research questions the work exposes.
Can one personal relationship preserve intent and evidence across agents while reducing what the person must carry?
FDE & AI Engineer Intern
AtlanWorked where models became production systems with context, tools, credentials, integrations, deployment, and real failure modes.
What does it take to know that an agent's finished run actually resolved the human outcome?
MTS Intern · LLM Pre-training
Soket AI Labs · Project EKAWorked below the model interface on multilingual curation and reproducible experimentation for an IndiaAI Mission-backed sparse-MoE program.
How much of model capability is decided by data and evaluation before architecture receives the credit?
MIRAI-Setu Participant
India–Japan ExchangeSelected among 40 students from India by Japan’s Ministry of Foreign Affairs for a month-long India–Japan exchange, including a 15-day technology internship and meetings with public and technology leaders in Fukuoka.
What makes a technical system worthy of trust over decades rather than impressive for a launch?
AI Engineer Intern
OpenFn (C4GT)NLP-to-workflow pipeline for government-tech data integrations. Selected under the Code for GovTech national open-source program.
How do language models become dependable components inside systems with explicit schemas and consequences?
HackByte Lead Organiser
The Programming Club · IIITDM JabalpurHelped grow HackByte from an internal college event to 5,154 registrations across three years. As a Programming Club core member and mentor, I also started ML Summer School and Winter of ML for structured projects and peer learning.
How do you make technical depth inviting without lowering the standard?
B.Tech — Smart Manufacturing
IIITDM JabalpurCompleted a B.Tech in Smart Manufacturing as branch topper with an 8.5 CPI, studying AI/ML, mechatronics, industrial systems, and design at the boundary between software intelligence and the physical world.
What changes when intelligent software must perceive and act in a noisy physical world?
Things I build with.
Agent systems
- Harness Architecture
- Context Composition
- Tool & Permission Design
- Durable Execution
- Recovery & Observability
- Outcome Evaluation
Models and data
- Pretraining Data
- Dataset Curation
- Multilingual Systems
- Mixture-of-Experts
- PyTorch
- Model Evaluation
Physical and cyber-physical systems
- Industrial IoT
- Sensing
- Edge Inference
- Automation
- Control Systems
- Manufacturing Systems
Currently exploring: Persistent agents, long-horizon evaluation, user-owned memory, and physical AI
Let’s Build Something
Have a question, a disagreement, or a thread worth following? Write to me.
Nagpur, India