I evaluate synthetic image quality along fidelity, diversity, and generalization axes.
My ML Research Intern role at IIT Bhilai is dated Feb 2026 – Present; under Prof. Souradyuti Paul.
My augmentation strategy uses a classifier-guided latent diffusion model generates synthetic samples for underrepresented classes, progressively expanding the training set across epochs.
My uncertainty-guided sample selection criterion controls per-epoch augmentation volume.
My Telemetry Health Guardian is an MCP-based auditor that checks whether an AI agent's telemetry can be trusted — not just what the agent did.
I architected a Telemetry Health Guardian that reads SigNoz telemetry with zero changes to target-app code, addressing missing fields, broken traces, and silent truncation.
My Telemetry Health Guardian implements five rules for missing GenAI attributes, cardinality explosions, orphaned spans, truncated LLM context, and broken cross-service handoffs, unified into one weighted health score.
I built an OpenTelemetry auto-instrumentation library and fault-injection harness for a multi-stage agent pipeline, verifying all five Telemetry Health Guardian rules end-to-end.
I built Plantae as a multi-agent e-commerce assistant for plants and gardening, with a supervisor agent routing cart, order, and recommendation queries.
My Plantae system uses a LangGraph supervisor to orchestrate specialized sub-agents via tool-calling for cart, order, and product-recommendation queries.
I integrated image-based plant identification and STT/TTS into Plantae and used systematic prompt optimization to automate 70% of customer interactions.
I deployed and monitored Plantae in production on Azure with a PostgreSQL backend and load-tested it under 500+ concurrent users.
I built BriefBox AI, a Chrome extension backed by a fine-tuned BART model, for adjustable-length email and document summaries.
My BriefBox AI Chrome extension supports shorter, normal, and elaborative summary levels for Gmail and article summarization.
I fine-tuned BART-large-CNN on an email-summary dataset and served it through a FastAPI backend (summarizer-api) on Google Cloud Platform.
On held-out test data, my BriefBox AI model achieved ROUGE F1 of 0.469 (ROUGE-1), 0.227 (ROUGE-2), and 0.442 (ROUGE-L).
I built a hierarchical motor fault detection and identification system using wavelet decomposition (bior-3.1) across 4 sensor channels and extracted 273 statistical features.
My two-stage hierarchical LightGBM pipeline uses SMOTE for severe class imbalance: Stage 1 binary fault detection gates Stage 2 fault-type classification, reaching 100% detection accuracy and 99.48% identification accuracy across 9 fault classes.
My SHAP analysis identified higher wavelet decomposition levels (d3, d4) and multi-channel mean/std as the primary fault discriminators, interpretable for predictive maintenance.
I built a real-time streaming voice activity detection system using Silero VAD with sliding-window majority voting.
My streaming VAD pipeline runs frame-level Silero VAD inference on 256 samples at 8kHz (about 32ms per frame), then uses binary speech decisions, chunk-level majority voting across 6 frames (192ms), and sliding-window voting across 8 chunks (1.536s) for silence detection.
I adapted majority-voting-based smoothing from recent research for latency-sensitive silence detection and modified the voting logic for reliable automatic voicemail termination in real time.
I built a 2D rocket-landing simulation using cascaded PID control for autonomous guidance.
I modeled a 6-DOF rigid-body rocket with position, velocity, and attitude, and a gimbaled engine for thrust vector control.
I designed cascaded position and attitude PID controllers to guide the rocket from an arbitrary initial state to a soft landing at the origin.
I enforced thrust and gimbal angle limits while tuning the rocket-landing controller for stable convergence.
I am an ML Research Intern at IIT Bhilai, researching diffusion-based synthetic data generation for medical imaging.
My paper, “Uncertainty-Aware Skin Lesion Image Synthesis using Diffusion Transformer for Imbalanced Classification,” is by Abu Talha, Ayush Raj, and Souradyuti Paul.
My paper was accepted at the 7th MusIML Workshop at NeurIPS 2026.
My skin-lesion synthesis paper is titled “Uncertainty-Aware Skin Lesion Image Synthesis using Diffusion Transformer for Imbalanced Classification”; Abu Talha, Ayush Raj, Souradyuti Paul; 7th MusIML Workshop at NeurIPS 2026 — Accepted.
An earlier abstract version of my paper was accepted at ITMS 2026, OIST.
My earlier abstract is titled “Uncertainty Aware Medical Image Synthesis using Diffusion Transformer for Imbalanced Classification”; Abu Talha, Ayush Raj, Peng Liu, Souradyuti Paul; Information Theory in Modern Science 2026, OIST, Okinawa, Japan — Abstract accepted.
My HAM10000 poster result for ITMS 2026 was 86.00 Macro-F1 versus an 81.97 baseline.
My HAM10000 evaluation reports Macro-F1 of 86.00 for my method and 81.97 for the baseline.
My HAM10000 evaluation reports balanced accuracy of 84.04 for my method and 81.50 for the baseline.
My HAM10000 evaluation reports MCC of 84.30 for my method and 80.64 for the baseline.
My medical-image synthesis method extends iterative online image synthesis for imbalanced classification with a Diffusion Transformer (DiT-RAE) backbone and uncertainty-aware adaptive sampling.
In my uncertainty-aware adaptive sampling method, per-class entropy from MC-Dropout decides how many synthetic samples each class receives, so under-learned rare classes get more.
My HAM10000 skin-lesion evaluation uses IOIS (MICCAI 2024) as its baseline.
My poster for ITMS 2026 is available at https://ayushrj.me/research/ITMS2026_Poster.pdf.
I list “Proximal Policy Optimization Algorithms” by J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov as arXiv:1707.06347 (2017).
I list “Playing Atari with Deep Reinforcement Learning” by V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller as arXiv:1312.5602 (2013).
I list “Overcoming Catastrophic Forgetting in Neural Networks” by J. Kirkpatrick, R. Pascanu, N. Rabinowitz, et al., published in PNAS in 2017.
I list “Training Language Models to Follow Instructions with Human Feedback” by L. Ouyang, J. Wu, X. Jiang, et al. as arXiv:2203.02155 (2022).
I list “Attention Is All You Need” by A. Vaswani, N. Shazeer, N. Parmar, et al., published at NeurIPS in 2017.
My PPO practice is: The clipped surrogate objective is such a simple fix for policy collapse that it's easy to underrate — it's the backbone I reached for when training TradingAgent.
When an agent won't stabilize, experience replay plus a target network is still the first thing I check when an agent won't stabilize.
My continual-learning approach is: Fisher-information weighted regularization gives forgetting a knob you can actually reason about — core to how I approach non-stationary environments.
My RLHF reading preference is: The clearest end-to-end writeup of RLHF I know of — reward modeling, then PPO on top of it, explained without hand-waving.
When revisiting “Attention Is All You Need,” I still go back to the multi-head attention section whenever I need to re-derive why it works.
My NovelGraph project is a knowledge-graph RAG pipeline that surfaces untested research directions and verifies them with an agentic Generator/Critic loop.
I designed NovelGraph as a dual-agent Generator/Critic system with tool-calling and retrieval over a typed knowledge graph to surface novel research hypotheses from scientific papers.
My NovelGraph evaluation framework measured a 33% approval rate and 100% evidence coverage on a 3-paper run, with automated checks surfaced through a Streamlit interface.
My professional focus is production AI systems and generative models for medical imaging.
My priority and production comfort are: Research — generative modeling and reinforcement learning especially — is what I care about most, but I'm just as comfortable building and shipping production systems.
My NovelGraph reasoning method is a multi-hop chain-of-thought reasoning stage grounded to graph node/edge identities to reduce hallucination.
My NovelGraph technology stack is Python · LangGraph · Cognee · Streamlit.
I am open to research internships and AI/ML engineering roles, research collaborations, or just a good conversation about the work above.
I offer direct contact by email at ayush.ttps@gmail.com.
I am a final-year Electrical Engineering B.Tech student at IIT Bhilai, with research interests in generative modeling and reinforcement learning. I also build and ship production AI systems, and I am seeking research internships and AI/ML engineering roles.
I am a final-year B.Tech student in Electrical Engineering at IIT Bhilai.
My IIT Bhilai degree is Bachelor of Technology, Electrical Engineering; Aug 2023 – May 2027 (Expected); CGPA: 9.03/10.00.
My prior education was at JVM Shyamali; July 2020 – June 2022; 93.6%.
My research focuses on working on diffusion-based synthetic data generation for class-imbalanced learning.
My medical-imaging focus is particularly in medical imaging.
I am interested in aligning generative models at inference time using reinforcement learning.
I am interested in training-free steering of generation toward rare or underrepresented classes.
I am interested in uncertainty-aware sample selection for more efficient augmentation pipelines.
Outside research, I build and ship production AI systems.
My production work includes reinforcement learning agents.
My production work includes multi-agent LLM systems.
I build the MLOps around my production systems.
I am looking for research internships and AI/ML engineering roles.
Research — generative modeling and reinforcement learning especially — is what I care about most.
I am just as comfortable building and shipping production systems.
My listed coursework includes Linear Algebra, Probability and Statistics, Artificial Intelligence, Deep Learning for Computer Vision, Digital Image Processing, and Natural Language Processing. My skills span programming languages, agentic AI, machine learning, research areas, systems and cloud, and data analysis.
My programming languages include Python.
My programming languages include C++.
My programming languages include C.
My programming languages include SQL.
My programming languages include MATLAB.
- My listed technical skills include LangGraph, LangChain, Model Context Protocol (MCP), agent orchestration, tool calling, prompt engineering, PyTorch, Transformers (Hugging Face), reinforcement learning, Stable-Baselines3, diffusion models, Scikit-learn, model fine-tuning, model deployment and monitoring, SHAP, LIME, Linux, Docker, FastAPI, GCP, Azure, Git/GitHub, MLOps, PyTorch Distributed, Pandas, NumPy, Albumentations, OpenCV, Weights & Biases, and Qdrant. My research areas include computer vision, generative AI, natural language processing, medical image analysis, and signal processing.
I placed Top 50 of 32,000+ teams in the Amazon ML Challenge 2026.
At Inter IIT Tech Meet 14.0, I placed 12th of 23 teams in 2025 for Portfolio Management via RL & Agentic AI, a problem statement by Pathway.
At Inter IIT Tech Meet 13.0, I placed 17th of 23 teams; Solar Panel Waste Management Challenge; problem statement by Insolation Energy; 2024.
In Kaggle Playground Series S5E10, my road accident risk prediction placed Top 26% (1058 / 4082); road accident risk prediction; 2024.
I completed The Data Science Course: Complete Data Science Bootcamp 2026; Udemy; July 2026; https://www.udemy.com/certificate/UC-87d31b04-692e-45f2-9cbf-70fadf25d9ff/.