Ashok Prasad Neupane

ML Researcher & AI Engineer · Kathmandu, Nepal

I work on reinforcement learning for generative models and LLM agents, controllable 3D generation, and efficient streaming vision-language models. I'm a founding AI engineer (1 of 2) at Elephant Intelligence, building browser-use agents, and first/co-first author on papers at CVPR 2026 Workshop and under review at ICLR, WACV, and TMLR.

B.E. in Computer Engineering, Pulchowk Campus (IOE, Tribhuvan University) — ranked 47 of 15,000+ in the national entrance exam.

Email Google Scholar alphaXiv ORCID Semantic Scholar Hugging Face GitHub LinkedIn CV (Research) Resume (Industry)

Publications

iARCS: Iterative Agentic RL for Controllable 3D Scene Generation

Saugat Adhikari*, Ashok Prasad Neupane*, Pramish Paudel, Ajad Chhatkuli, Danda Pani Paudel

CVPR 2026 Workshop · extended version under review at WACV

Agentic RL that adapts a pretrained 3D scene diffusion model to natural-language constraints using LLM-written reward programs refined in a reflection loop. Task success 4.5% → 73.8% (robot reachability) over MiDiffusion; 12K generated scenes released.

Project Page arXiv Code Dataset HF Paper

Topics: controllable 3D scene generation · indoor scene synthesis · reinforcement learning for diffusion models (DDPO) · Eureka-style LLM reward design · embodied AI

Foresight: Planning Future Perception in Streaming VLMs without Retraining

Ashok Prasad Neupane* (corresponding), Dipan Bartaula*, Ankit Belbase, Saugat Adhikari, Samip Ghimire, Saroj Poudel, Binod Bhattarai, Danda Pani Paudel

Under review at ICLR

Training-free dual-stream architecture: two Siamese LLMs sharing weights and KV cache — one ingests frames, the other plans when to reason and what to check. With a frozen Qwen3-VL-8B, beats the strongest trained baseline by 9.5% on OmniPro Online; +6.7 on StreamingBench and +15.4 on OVO-Bench.

Project Page arXiv Code HF Paper

Topics: streaming video LLM · streaming VLM · proactive real-time video understanding · training-free · asynchronous inference

iADD: Improving Alignment and Diversity in Diffusion Policy Optimization

Ashok Prasad Neupane*, Saugat Adhikari*, Pramish Paudel, Ajad Chhatkuli, Danda Pani Paudel

Under review at TMLR

Shows late-timestep-only RL updates harm diversity and proposes incremental-timestep training with Feynman–Kac particle guidance. Rare-prompt success 24.8% → 66.9%; reward–diversity AUC 53.8% → 77.7% vs. DDPO.

Project Page arXiv Code HF Paper

Topics: diffusion policy optimization · RL fine-tuning of diffusion models (RLHF) · DDPO · Feynman-Kac steering · mode collapse and reward hacking

* equal contribution

Experience

Elephant Intelligence — Founding AI Engineer (1 of 2)

Jun 2026 – Present · San Francisco (Remote)

Building an agent harness for browser-use agents for web form filling and navigation; own the core product and evaluation stack end to end.
Built an in-house benchmark on BrowserGym; drove task success from ~20% to 100%, ~2× faster and 25%+ cheaper than the closest competitor.
Shipped as a desktop app, REST API, and browser extension on AWS.
Built an LLM routing system beating recent published methods by 10%; cut per-task cost 4–5× with continual learning.

Sceneverse — Machine Learning Engineer

Jan 2025 – May 2025 · Remote

Built 3D generative AI training and inference workflows for controllable indoor scene synthesis.
Developed synthetic data pipelines for photorealistic 3D asset rendering, inspired by Apple's Hypersim.
Trained and evaluated 3D scene foundation models on multi-GPU compute.

Freelance AI Engineer

Jan 2024 – Present · Remote

End-to-end text-to-video pipeline: LLM scripts, diffusion lip sync, audio-visual alignment, voice cloning.
RAG-based grading agent for schools in Kathmandu Valley.
Lightweight VLM deepfake detector: 98.7% accuracy on seen, 87.3% on unseen generators.
AI Chrome extension for QA automation that maps dynamic DOM IDs to stable selectors.

Fusemachines — AI Fellow

Dec 2022 – Feb 2024 · Kathmandu

15-month fellowship in ML, deep learning, CV, NLP, and production ML.

Teaching

Mid-Valley International College — Instructor

Kathmandu

Taught undergraduate courses in Data Science and AI & Machine Learning.

Skill Shikshya — AI & Programming Instructor

Oct 2024 – Jan 2025 · Kathmandu

255+ hours of Python, prompt engineering, agentic AI, and RAG training.

Projects

▶

Human Action Recognition (with Khalti)

VideoMAE model with 95%+ accuracy on Nepali video data; log-linear sequence scaling for long videos.

Demo Code

▶

Adaptive Traffic Light System

ESP32/Arduino controller with real-time CV vehicle detection driving adaptive signal timing.

Demo

▶

Olympic Sentiment Analysis

FastAPI + React app serving a fine-tuned BERT model for real-time sentiment of Olympic posts. GeeksForGeeks Geek-A-Thon.

Demo Code

Fashion Virtual Try-On

Distilled diffusion models for 20× faster try-on without fidelity loss. KU Hackfest AI Track Winner; Hult Prize finalist.

Code

Awards & Honors

CVPR 2026 Workshop — iARCS presented at the Synthetic Data for Computer Vision Workshop, Denver
KU Hackfest — AI Track Winner (Nov 2024)
Hackademia 2 — Best AI Product (Sep 2024); Hackademia — Best Presentation (Dec 2023)
Samsung Innovation Campus — Top 30 nationally (Coding & Programming; Big Data Computing)
Generative AI & LLMs Course — Top 20, Pulchowk Campus
Geek-A-Thon (Aug 2023) · Engen Hacks (Oct 2023) · Civil Engineering Hackathon (Dec 2024) · Leo Club Ideathon (Dec 2024)
BE Model Winner, Pulchowk Campus (Jan 2024) · All Nepal Civil Engineering Model Competition Winner (May 2025)

Presenting iARCS at CVPR 2026 — CVPR 2026, Denver