A language-model agent that simultaneously modifies task-specific scaffolding and model weights. Achieves 25.1% over prior SOTA on LawBench and 12.4% faster GPU kernels.
ICML 2026
4 papers accepted at ICML 2026, held in Seoul, South Korea - including SIA, our self-improving AI system.
At the ICML 2026 venue, Seoul, South Korea
View all photosCurrent Research Focus
My main focus right now: small language models (SLMs). Everyone should own their own AI - a model small enough to live on your hardware, cheap enough to run continuously, and smart enough to be worth talking to. The frontier isn't just bigger models; it's models that are good enough, running everywhere.
- On-device small language models: Quantized inference (4-bit weights, fused kernels), KV-cache-efficient decoding, and latency budgets tight enough to run a capable model on a phone or laptop.
- Self-improvement at small scale: Distillation from frontier models, continual and test-time adaptation, and lightweight fine-tuning so a small model gets better on your data, on your device.
Why small models are a research problem
L(N, D) = A · N−α + B · D−β + E0
What this tells me
The scaling law says: a model's loss falls as a power law of parameters N and data D, but each doubling buys less and less, and you can't go below the irreducible floor E0. For an SLM this is brutal - at 1-3B parameters you slam into diminishing returns fast. What that tells me: to make small models genuinely useful you can't brute-force scale your way out. You have to beat the curve with better architecture, better data, and self-improvement that compounds outside a data center.
Research
View all papers →Autonomous self-refinement via evolving agent structure and test-time reinforcement learning. +16pp on LawBench, -19% GPU kernel runtime.
Addresses the cost/reliability tradeoff in proxy evaluations. MLEvolve achieved SOTA MAE of 0.1354 on MLE-bench within 12 hours.
A two-agent system pairing a Scientist with an advisor that can only ask questions. Improved Kaggle test scores on 4/5 MLE-bench tasks with a mean increase of ~56%.
A benchmark for evaluating whether an AI agent can modify another agent to improve it, covering meta-improvement and self-improvement scenarios.
Experience
Hexo LabsSenior Research Scientist
Leading AI research and high-impact programs end-to-end, from architecture and novel research to production delivery.
- 4 papers @ ICML 2026
- Co-authored & co-built SIA
- Led tech across multiple AI projects
- Owned end-to-end delivery for client programs
AI CallerAI / Backend Engineer
Built a real-time AI calling system before audio-to-audio models existed. Engineered low-latency voice pipelines from scratch and took it from zero to revenue.
- $2,000+ MRR within 2 months
- Pre audio-to-audio era
- Ultra-low latency voice AI
- Sole engineer on critical infrastructure
Who I've Worked With
Organizations I've worked with across AI research, engineering, and product development.
DPIIT
IP India
Bito
Soliton
Atomicwork
Dashtoon