Current
Mar 2026 — Present
AI Research Scientist
Wix
Making Websites beatiful — not like this one ):
About
Profile
I'm an AI Research Scientist at Wix, working across computer vision and NLP with a focus on multimodal models and text-to-image / text-to-video diffusion. My work bridges research and engineering: building data and training pipelines, running rigorous evaluations, and getting models to run on a single GPU — and, when needed, on the edge.
Experience
2023 — Present
Timeline
-
Current
Mar 2026 — Present
-
Jan 2025 — Jan 2026
Computer Vision Research Engineer
Reality Defender
Developed and optimized deep-learning computer-vision systems for deepfake and synthetic-media detection across image, video, and video+audio. Multi-model architectures combining specialist detectors with shared backbones for production-scale inference.
-
Dec 2024 — Apr 2025
Computer Vision Research Engineer
LuckyLab · Freelance
Edge-optimized segmentation and detection for production in a few-shot domain. Converted research models to production graphs (JAX, TensorRT) with quality gates using pruning and quantization.
-
Jun 2024 — Jan 2025
Tel Aviv
Deep Learning Research Engineer
NLPearl
Compact language models for multi-task outputs, audio tokenizers and LLMs for audio tasks, and real-time conversational pause detection with fine-tuned LLMs. LoRA and multi-stage training under tight latency budgets.
-
May 2023 — Jun 2024
Rehovot
Computer Vision & Deep Learning Research Engineer
Pashoot Robotics
Zero/few-shot improvements with SAM, YOLO-World, Grounding DINO, and CLIP. Multi-object tracking and 6-DoF pose estimation using synthetic data and domain randomization, plus 3D reconstruction (NeRF, Gaussian Splatting, image-to-3D) for simulation.
Featured
In progress
Latest research
Current projects — drag the slider to compare base output to the ES-trained LoRA, and watch an agent play MapleStory on its own.
Evolution Strategies
RL
Text-to-image
Post-training
ES-EGGROLL — Post-training Text-to-Image with Evolution Strategies
An EGGROLL-style implementation of low-rank Evolution Strategies from "Evolution Strategies at the Hyperscale" (Sarkar et al., 2025), adapted for text-to-image alignment post-training. The base generator stays frozen, and we optimize only a LoRA adapter using black-box reward signals (e.g., PickScore as the objective), with CLIP / aesthetic / no-artifacts as diagnostics. This enables fast iteration on alignment objectives at near inference throughput, without diffusion backprop.
Benchmark snapshot · PartiPrompts (overall means)
One image per prompt · same seeds across models · deltas vs. SanaOneStep_Base.
| Model | aesthetic ↑ | CLIP text ↑ | no artifacts ↑ | PickScore ↑ |
|---|---|---|---|---|
| SanaOneStep_Base | 0.5978 | 0.6592 | 0.3859 | 22.3220 |
| SanaOneStep_eggroll | 0.5975 | +0.00190.6611 | +0.00400.3899 | +0.179322.5013 |
| SanaTwoStep_Base | 0.5965 | 0.6614 | 0.3926 | 22.8059 |
Prompt
Use ← / → to switch examples
Example
1 / 1
← Base
ES-LoRA →
⇆
Drag horizontally (or use the slider) to reveal more of the ES-trained LoRA result.
Agents
Behavior cloning
Vision encoder
LLM policy
Maple LLM — An Agent That Plays MapleStory
Teaching a small LLM to play MapleStory (on a private server) by behavior cloning my own gameplay. At play time the screen is the only input: a DINOv3-based encoder compresses every frame into a single 512-d token, and an LFM2.5-230M policy — the same tiny model that chats in the Playground below — reads the token stream plus action history and predicts the next held-key combo at 10 fps.
Bootstrapped supervision
- Record my own play — 18 sessions · 5.6 h · 301k frames, with a synced keyboard log.
- Label just 267 frames with a SAM-2.1-assisted tool (3.9k boxes, 7 classes).
- Detect — a YOLO11m trained on those labels pseudo-labels every frame.
- Compress — the encoder distills each frame into one 512-d token, forced to carry game state by aux losses on the pseudo-labels.
- Act — LFM2.5-230M SFT over 28 held-key combos.
Hardest part: copycat. Held keys persist, so ~83% of targets just repeat the previous action — the model happily ignores the screen. Countermeasures: action-history dropout and transition-weighted loss; the metric that matters is transition accuracy.
Selected work
Projects
Research and engineering work, filterable by domain.
Tom & Jerry — SANA-Video LoRA
Fine-tuning SANA-Video 2B with LoRA to generate Tom & Jerry-style animations at 224×224 — from class-only LoRA to scene-aware text conditioning, all on a single consumer GPU.
T2V LoRA
VLM + vLLM
Diffusion video
Read more
Versioned training roadmap
- V1 — Class-only LoRA @ 224×224. Freeze SANA-Video 2B and train LoRA on all linear layers using a single class prompt over ~16k cached 5s clips (81-frame latents) with flow-matching.
- V2 — Scene-aware LoRA with VLM labels. Use Qwen3-VL via vLLM to generate structured scene descriptions (ENVIRONMENT / CHARACTERS / PROPS / ACTION / CAMERA), then condition SANA-Video on per-clip text prompts.
V1 checkpoint progression
Same seed and class prompt; only the LoRA checkpoint changes.
Base model: off-style 224×224 sample before LoRA fine-tuning.
Future directions
Next: distillation, pruning, and quantization for deployment, plus more systematic evaluation (identity consistency, motion quality) and RL-style post-training to better align generations with scene descriptions.
Sana Simplified — Image & Zoom Control
Research playground around Sana 1.5 and Sana Sprint: ControlNet implementation and a SigLIP-driven zoom controller that smoothly changes camera distance while keeping identity and style consistent.
T2I
Diffusion
ControlNet
3D rendering
Read more
Method overview
A zero-zoom reference image is encoded with SigLIP into an object token, while a scalar zoom value z ∈ [0, 1] is mapped through a small MLP to a zoom token. Both tokens are appended to the cached text encoding, and the Sana 1.5 transformer is fine-tuned with LoRA.
Reference & zoom sweep
Kokoro — Tiny TTS, Big Voices
Small-footprint TTS with Kokoro-82M: fast mixture-of-voices coefficients, full-embedding optimization, and visual tooling (trajectory visualizer + export).
TTS
Embedding opt
Visualizer
Read more
Highlights
- Mixture-of-voices coefficient optimizer with temperature-annealed softmax.
- Optimizing voice embedding and LoRA in few-shot scenarios.
- GUI visualizer: circular mixer + per-voice bars + MP4 export.
Token-Budget-Aware Reasoning for VLMs
Built on Token-Budget-Aware LLM Reasoning, extended to multimodal setups by combining a frozen SigLIP image encoder with a LoRA-tuned LLM that predicts a reasoning budget prior to decoding.
CoT
Multimodal
Paper impl.
Read more
Summary
The original paper asks whether we can control chain-of-thought with an adaptive token budget. This variant predicts the budget for multimodal inputs and constrains decoding accordingly.
Deep dive
- Architecture: frozen SigLIP + budget head fused with LLM states.
- Training: LoRA fine-tuning with oracle CoT lengths; KL regularization.
- Evaluation: accuracy vs. average tokens; ablations on freezing / head depth.
PopYou2 — VAR Text-to-Image
Adapted Visual AutoRegressive (VAR) for Funko Pop! generation with a custom "doll" embedding and a SigLIP→text adapter, enabling I→I and T→I paths.
VAR
T2I
Autoregressive
Few-Shot SAM + LoRA
Adapted SAM and efficient variants (EdgeSAM, TinySAM) with LoRA for class-aware few-shot segmentation; improvements over PerSAM and edge deployment.
Few-shot
LoRA
Pruning
Quantization
Edge
Read more
Deep dive
LoRA-adapted SAM variants with compression for edge deployment, with experimentation in extreme few-shot scenarios.
CelebrityLook — Mobile Face Transform
On-device face stylization with a distilled StyleGAN2 and CLIP-to-StyleGAN latent alignment; ~30 fps on modern phones.
Edge
GAN
CoreML
Latent alignment
PopYou — FastGAN + CLIP
Multi-stage pipeline with FastGAN and CLIP inversion for promptable Funko-style synthesis.
GAN
CLIP inversion
Synthetic data
MusicGen — Genre LoRA
LoRA-adapts MusicGen for genre-specific generation while keeping prompt controllability.
LoRA
Music generation
KoalaReadingAI — Papers as Podcast
Pipeline that turns AI research papers into short audio episodes published to Spotify and YouTube.
TTS
Pipeline
Playground
Fully playable
Toys that learn
Hands-on ML — drive an optimizer, out-judge a generator, compose Funko Pops with my VAR model, and chat with a tiny LLM that lives in your browser. Rumor has it not everything here is listed.
Descent — you are the optimizer
SGD with momentum, played by hand. Pick a learning rate, roll the ball to the flagged global minimum — local minima are traps, and too much lr ends in NaN.
Interactive
SGD + momentum
score = epochs
Real or Generated?
The deepfake-detection problem, reduced to its purest form: ML paper titles. Half are real published research. Half were hallucinated by a language model for this site.
Interactive
Discriminator: you
streak counter
Funko Forge — PopYou2, live
Real outputs from my VAR text-to-image model — compose the prompt, meet the figure. 36 combinations, zero GPUs harmed.
A Funko Pop figure of
styled as a
performing
.
rumor: one combination is legendary ⚔️
Interactive
VAR
Autoregressive T2I
contacting the hub…
Tiny Amit — a 230M-parameter me
LiquidAI LFM2.5-230M, quantized and running entirely in your browser — no server, no API. It has read a short note about me, so ask it anything. At 230M parameters it is fast, charming, and confidently wrong on schedule.
WebGPU preferred · WASM fallback · answers ≠ facts
Interactive
LFM2.5-230M
On-device inference
Reading list
Updated continuously
Recent papers
Auto-synced from my Zotero library — papers I'm currently working through.
Contact
Get in touch
Open to research collaborations, side projects, and interesting problems.