Ashwani Bhat

01. model card_

I'm a deep learning engineer and applied scientist working at the intersection of LLMs, NLP and multimodal ML. At Amazon I design systems that moderate and label content across global marketplaces — policy-aware retrieve-then-label pipelines, multi-agent label refinement, LLM-as-a-Judge evaluation councils, and cascaded inference architectures that keep latency and cost down.

Before that I researched NLP at IIT Kanpur (multimodal emotion recognition, adversarial text generation, sign-language corpora) and explored quantum ML, model interpretability and adversarial robustness at MathWorks.

I like problems where research meets production: scarce labels, shifting policies, many languages, and a serving bill that needs to shrink.

ashwani@gpu-rig: ~/career

02. training history_

Each role, logged like the training run it was. Loss only went down.

e3

Applied Scientist II · Amazon

Oct 2024 — present · Bengaluru

Built a policy-aware retrieve-then-label framework with multi-agent iterative label refinement — labels stay compliant through policy updates 2–3×/month, with no retraining.
Deployed a vLLM Qwen2-VL endpoint with multi-LoRA, consolidating multiple endpoints into one and cutting serving cost by 70%.
Created an LLM-as-a-Judge evaluation framework (QA / RAG / classification) with a multi-LLM “council” that lifted labeling precision by 10%, plus a trace-visualization UI.
Designed a cascaded inference architecture for legacy ad-policy classifiers to cut cost and end-to-end latency.
Built a synthetic data pipeline with Qwen-Image to fix scarce, imbalanced labels and boost downstream quality.

vLLMmulti-LoRAQwen2-VLRAGmulti-agentLLM-as-a-Judge

e2

Applied Scientist I · Amazon

Jul 2022 — Sep 2024 · Bengaluru

Shipped content-moderation classifiers for data-scarce locales (Spain, Japan) using weak supervision over XLM-RoBERTa / BERT with custom architectural tweaks.
Built multimodal labeling models (FLAVA) for US/UK/CA marketplaces, raising automation rates.
Designed an automated hard-sample mining algorithm for ASIN categorization — 70–80% less manual effort, 5–10% faster model development.
Led operational-excellence initiatives and served as Bar Raiser for multiple model launches.

XLM-RoBERTaFLAVAweak supervisionmultimodalhard-sample mining

e1

Application Support Engineer · MathWorks

Aug 2021 — Jun 2022 · Bengaluru

Trained linear regression via adiabatic quantum computing — experiments on the D-Wave 2000Q QPU.
Implemented interpretability methods (Occlusion, LIME, Grad-CAM) in MATLAB's Quantization Module.
Added adversarial-attack support (FGSM, BIM) and evaluated robustness of quantized models under attack.

D-WaveGrad-CAMLIMEFGSM / BIMquantization

03. proof of work_

A multilayer perceptron, written from scratch in vanilla JavaScript — forward pass, backprop, the lot. It is training in your browser right now. Click the canvas to add cyan points, shift-click for violet, and watch the decision boundary chase them.

dataset

hidden layers 2

neurons / layer 8

learning rate

activation

epoch 0 loss —

architecture

2 → 6 → 6 → 1

04. vibe projects_

A side-quest log of small, weird, hands-on ML experiments — every one of them runs and trains live in your browser, most with zero libraries. Three of the current residents:

D16 · agentic systems

🔧 Harness Lab

The model is fixed; tune the harness around it. Budgets, verifiers, observation shaping, typed errors — same model, 34% → 92% of tickets solved.

play ↗ D17 · evaluation design

⚖️ Judge & Jury

An LLM-as-a-Judge protocol that scores 74% agreement — and still crowns the wrong model. Bias measured by counterfactual.

play ↗ D7 · world model

💤 Pong Dreams

A network watches Pong until it learns the physics — then you play inside its hallucination as it drifts.

play ↗

🧪 enter the lab — all vibe projects eight live demos and counting · new experiments land regularly

05. interview blogs_

Field manuals for senior ML interviews — long-form study guides with graded question banks, levelling rubrics, and case studies worked end to end with numbers.

field manual · agents & tool use

📚 all interview blogs

06. peer-reviewed checkpoints_

Research published at top-tier NLP and ML venues.

NAACL 2022 · main

07. nvidia-smi, but for me_

+-----------------------------------------------------------------------+
| ASHWANI-SMI 5.0      Driver Version: chai_v3      CUDA Version: 12.x  |
|--------------------------------+------------------+-------------------|
|  SKILL                         |  UTIL            |  MEMORY-USAGE     |
|================================+==================+===================|

+-----------------------------------------------------------------------+
|  Processes:  vLLM serving · multi-agent labeling · LLM-as-a-Judge     |
|              weak supervision · synthetic data gen · hard-neg mining  |
+-----------------------------------------------------------------------+

languages

PythonC / C++MATLABBash

deep learning

PyTorchTensorFlowKerasScikit-LearnOpenCV

llm & genai

vLLMLoRARAGMulti-agent systemsLLM evaluationQwen-VL

infra & tools

AWSGitVimLaTeXUbuntuMySQL

08. pre-training_

2021

M.Tech · Computer Science & Engineering

Indian Institute of Technology, Kanpur

val_score: 8.60 / 10

2018

B.Tech · Computer Science & Engineering

National Institute of Technology, Srinagar

val_score: 9.17 / 10 · dept_rank: 1 🏆

09. backprop to me_

Let's build something intelligent.

Open to collaborations, research chats, and hard ML problems.

✉ ashwanibhat44@gmail.com

github linkedin kaggle twitter