Final-year B.Tech CSE (AI & ML) at Woxsen University, Hyderabad. I build deep learning systems end to end — data, training, honest evaluation, and the part where somebody else can actually run it.
Currently building atlas — a local-first AI companion that diagnoses and explains your machine without sending a byte of it anywhere.
| Project | What it does | Where it got to |
|---|---|---|
| AI-Generated Image Detection | EfficientNet-B0 classifier for DALL·E / Midjourney / SDXL output, with Grad-CAM to show what the model keyed on | 98.5% on held-out test images, trained on 120k |
| atlas | Local-first, privacy-first assistant that diagnoses system problems and explains them in plain language. Runs fully on-device — no telemetry, no cloud | In active development |
| Continual learning with DCE-CVAE | Class-incremental learning on Split-MNIST via pure latent replay — four-tier memory hierarchy, per-task orthogonal subspaces, latent merging | ≈0.66 avg accuracy over 5 sequential tasks vs ≈0.20 for naive fine-tuning |
| symptoms-disease-reasoning | Diagnostic reasoning engine: rule-based scoring, ML ensemble priors, Tree-of-Thoughts questioning, and RAG explanations grounded in PubMed | Research prototype, interactive CLI + API |
| autostream-ai-agent | LangGraph agent that turns a conversation into a qualified lead — passive field extraction, confidence-thresholded intent routing, RAG-grounded replies | Deployable to WhatsApp via webhooks |
| multimodal-market-signal-fusion | Audio sentiment + TF-IDF text + technical indicators, fused with entropy-based adaptive gating and XGBoost | 0.627 acc / 0.616 F1 — under a leak-free chronological split, not a shuffled one |
Python · PyTorch · TensorFlow · scikit-learn · Transformers · LangChain · LangGraph · XGBoost · OpenCV · FastAPI · Ollama · C++
Computer vision and object detection · continual learning · retrieval-augmented and agentic systems · on-device and privacy-preserving ML — and increasingly the product side of all of it: what to build and why, not just how.