About Valfinity

Built from research on AI agents

Valfinity is founded by Jayant Parashar, a PhD researcher at the University of Georgia who specializes in self-improving AI systems.

Research

Preprint, 2026 · under review

Vestrum: Improving Agent Harnesses by Adapting Their Verification, Structure and Memory

J. Parashar, E. F. Douglass, W. C. Bastian, S. M. Bhandarkar

Turns an agent's execution failures into targeted changes to how it verifies, retrieves, plans and remembers, without retraining the underlying model. Tested across five settings, from long-horizon tasks to cell-type annotation in spatial biology.

In practice: Your agent learns from its own mistakes and gets more reliable every week.

arXiv:2609.33822 ↗

Preprint, 2026

KACE: Knowledge-Adaptive Context Engineering for Mathematical Reasoning

J. Parashar, S. M. Bhandarkar

A self-reflective loop that organizes past reasoning into a knowledge base sorted by difficulty and domain, then retrieves guidance specific to each new problem. No weight updates needed.

In practice: Your agent reuses what worked before instead of starting from zero.

arXiv:2606.00532 ↗

IEEE CAI 2025 · peer-reviewed

SJ-ORPO: Alignment of LLMs Using Multi-Task Learning and Self-Judgment

J. Parashar, S. M. Bhandarkar

Trains a language model (Mistral 7B) to generate answers and judge its own outputs at the same time, combining preference alignment with self-evaluation.

In practice: Your agent checks its own work before handing it to you.

doi:10.1109/CAI64502.2025.00040 ↗

63.7% → 70.3%Terminal-Bench 4 Hard checks passed, vs. Claude Code, at 1.03× the costVestrum

47.6 → 59.8UltraHorizon score on held-out tasks, vs. the GAM harnessVestrum

51.8% → 62.2%AIME 2025 accuracy, vs. Best-of-5 at a comparable budgetKACE

Results reported in the papers above, on held-out tasks.

Founder

Jayant Parashar

Before I started working on agents, I spent years building, running and understanding the models that agents run on. My PhD in Computer Science at the University of Georgia covers LLM post-training, preference alignment, reward signals for reinforcement learning, and continual learning. Now, my vision is to build a self-improving AI system with agents, for your specific use-cases!

I started in theoretical physics at the University of Delhi and shifted to AI because I believe superintelligence will arrive before physics is unified, bringing an age of prosperity and abundance. Through Valfinity, I plan to realize that dream and put a self-improving AI system in the hands of business leaders and startups.

PyTorch · Hugging Face · verl · LangGraph

LinkedIn ↗GitHub ↗

Join the mission

I'm looking for a cofounder who prides themselves on being a fast learner and a patient engineer.

Want an agent like this?

parasharjjjayant@gmail.com