"Not just running
.fit(). Building the entire universe from scratch."
dashpreet = {
"name" : "Dashpreet Singh",
"alias" : "DZDasherKTB",
"college" : "IIT Jammu : B.Tech CSE (2024–2028)",
"domains" : ["Deep Learning", "Computer Vision", "NLP", "Agentic AI",
"Multimodal Systems", "Hardware Acceleration", "Cybersecurity",
"Geospatial AI", "Spatial Transcriptomics"],
"vibe" : "Build weird stuff. Break it. Rebuild it better.",
"currently" : "Researching unified multimodal architectures for bioinformatics",
"patents" : 2,
"open_to" : ["Research Collabs", "Internships", "Hackathons", "Building things"],
}| 🥇 Achievement | 🏢 Org | 📅 Year |
|---|---|---|
| 🛰️ ISRO BAH National Finalist (Top 10) | ISRO / Govt. of India | 2025 |
| 🤖 Inter-IIT Tech Meet 14.0 — Geospatial AI (Rank 12, 23 IITs) | IIT Patna | 2025 |
| 🛒 Flipkart GRID 7.0 Semi-Finalist | Flipkart | 2025 |
| 🔬 Shell AI Hackathon — Top 10 | Shell | 2025 |
| 🏅 ATF India — National Top 200 (200,000+ applicants) | ATF India | 2025 |
| 📚 Academic Excellence Award — Top 3 Branch Rank | IIT Jammu | 2025 |
| 📜 Patent Granted — Dual-Encoder Spatiotemporal Architecture | Govt. of India (App: 202511114061) | Nov 2025 |
| 📜 Patent Filed — Hybrid Multi-Modal Deep Embedding Recommendation System | Govt. of India (App: 202511098795) | Oct 2025 |
| 📜 Patent Filed — STAR Engine: Linear Memory Hardware Accelerator | IIT Jammu / Govt. of India (App: 202611043901) | Apr 2026 |
📄 Patent 1 — Dual-Encoder Spatiotemporal Architecture (GRANTED ✅)
App No: 202511114061 | Filed: Nov 2025 | Co-inventor: Dr. Iqbal Kaur
Title: Dual-Encoder Spatiotemporal Architecture with Multi-Modal Skip Fusion for Multi-Domain Environmental, Climatic and Agricultural Forecasting
A deep learning architecture that fuses static (Conv3D) and dynamic (ConvLSTM) tensors via a Cross-Attention Transformer (CAT) mechanism for spatio-temporal environmental prediction. Validated on MODIS, VIIRS, and NASA FIRMS satellite data.
📄 Patent 2 — Hybrid Multi-Modal Deep Embedding Recommendation System (FILED)
App No: 202511098795 | Filed: Oct 2025 | Published: Dec 2025
Title: Hybrid Multi-Modal Deep Embedding Recommendation System with Synthetic Data Generation and Fusion Learning
A recommendation architecture combining multi-modal embeddings, synthetic data generation pipelines, and fusion learning for scalable retrieval at industrial scale.
📄 Patent 3 — STAR Engine: Linear Memory Hardware Accelerator (FILED)
App No: 202611043901 | Filed: Apr 2026 | Applicant: IIT Jammu
Title: An Apparatus for Accelerating Reductionist Matrix Operations with Guaranteed Linear Memory Traffic
A dedicated hardware accelerator featuring a Streaming Tile-Autonomous Residency (STAR) Engine with a pipelined systolic scoring unit, online softmax, and probability-value multiplication — all on-chip. Reduces attention mechanism memory complexity from O(N²) to O(N). Demonstrated ~97x memory traffic reduction at N=8192 vs. baseline GPU multi-kernel.
🔬 Ongoing Research — The Tensor-Only Unit (TOU)
Advisor: Dr. Uma Satya Ranjan, IIT Jammu
Theoretical framework for post-GPU AI acceleration — designing a compute paradigm beyond GPU architecture, targeting tensor-native hardware execution.
🔬 Research Exploration — Omni-ST: Unified Instruction-Driven Multimodal Model for Spatial Transcriptomics
Researching a unified any-to-any multimodal framework for spatial transcriptomics, combining histology image encoders, gene expression adapters, spatial graph networks (GCN/GAT), and instruction-conditioned fusion transformers. Inspired by architectures like Kosmos-2, OpenFlamingo, STFlow, SpaGCN, and scVI.
Tasks: Image→Gene Expression | Gene→Cell Type | Spatial Domain ID | Region→Text Explanation
Inter-IIT Tech Meet 14.0 | Presented to 23 IITs at IIT Patna | Hosted on Modal
Built a multimodal geospatial agent "GeoChat" that decomposes natural language queries into executable geospatial pipelines.
- Stack: Qwen2-VLM, Qwen2.5-LLM, OWLv2, SAM, LangChain, Modal, Docker, OpenCV
- Orchestrated a 3-modality any-to-any pipeline: captioning → detection → segmentation
- LangChain workflow with Intent Parsing + Tiling Engine: VQA, Fact Manipulation, Spatial Grounding
- 85% intent parsing accuracy, 0.78 F1 score on spatial grounding tasks
ISRO Bharatiya Antariksh Hackathon — National Finalist (Top 10)
A dual-encoder architecture for forest fire prediction and spatial spread simulation.
- Stack: PyTorch, GEE, Conv3D, ConvLSTM, Cross-Attention Transformer, Cellular Automata
- Dual-Encoder: Static stream (Conv3D) + Dynamic stream (ConvLSTM) fused via CAT mechanism
- Cellular Automata for fire spread simulation
- Validated on MODIS, VIIRS, NASA FIRMS multi-source satellite data
- 9% improvement over SOTA fire prediction model in lead-time accuracy
Graph Neural Networks + Segment Anything Model for medical imaging
Combining GNN spatial reasoning with SAM's segmentation capabilities for medical use cases.
- Stack: Python, PyTorch, SAM, GNN
- Graph-augmented medical image segmentation pipeline
Production-grade VPN infrastructure inspired by Kevin Mitnick
- Full WireGuard VPN setup and management in Python
- Named after the legendary hacker — built with security-first principles
Adobe (XAI Track):
- 99.14% accuracy in GenAI content detection via Hybrid Model
- Explainable AI pipeline using GradCAM++
- LoRA/Fine-tuning benchmarks on Qwen2.5 (29k+ samples on VRSBench)
DevRev (Agentic AI Track):
- RAG-based Agent 007 using Planner-Filler-Validator workflow
- Dynamic tool-use orchestration with hallucination checks
Full-stack Text-to-Video Generation Platform
- Stack: FastAPI, Node.js, Next.js, Vercel, Supabase
- Async backend orchestrating GenAI models → short videos from text prompts (images + audio + captions)
Hybrid neuro-symbolic pipeline for intelligent document understanding
No high-level APIs. No shortcuts. Raw math → working code.
Deep Learning (PyTorch only — no Keras, no Lightning):
- Stable Diffusion (VAE + UNet + cross-attention) — optimized 16s → 0.8s/iter
- StyleGAN, DCGAN, Conditional GAN
- Transformers (BERT/GPT), full Encoder-Decoder translation (EN↔FR, EN↔DE, EN↔HI)
- R-CNN, Fast R-CNN, Faster R-CNN, YOLO
- U-Net, DeepLab V1→V3+
- Diffusion models (DDPM) — forward noise + reverse UNet denoising
- RNNs, LSTMs, GRUs — sequence generation, translation, prediction
Classical ML (NumPy only):
- XGBoost (Taylor expansion, gain scoring, GPU accel)
- Random Forest, AdaBoost, Gradient Boosting
- SVM (Hinge loss, Lagrange dual)
- KMeans, DBSCAN, Decision Trees
- Stacking, Voting, Blending ensembles
- KNN, Naive Bayes (Multinomial + Gaussian)
25+ real-world projects: Cancer/Heart/Stroke prediction, IMDB sentiment, Spam detection, MNIST, Titanic, Recommendation Systems, and more.
-181717?style=flat-square&logo=typescript){kind=icon}
- Auto-syncing portfolio with CRON-powered backend, SQL DB, SOLID architecture
- Full-stack TypeScript/React apps with Prisma + PostgreSQL
- Rate-limited backend mail systems, API-first modular design
- Phishing/social engineering analysis (educational research)
- WireGuard VPN infrastructure
- Project Mitnick — VPN + security tooling
-181717?style=flat-square&logo=github){kind=icon}
- Assembly language implementations
- Hardware accelerator design (STAR Engine patent)
- Custom backpropagation engines in raw PyTorch tensors
AI / ML / DL
Architectures & Models
-22D3EE?style=flat-square){kind=icon}
Web / Backend
DevOps / Infra
-5B21B6?style=for-the-badge){kind=icon}
Languages
| Platform | Handle | Status |
|---|---|---|
| 🟦 Codeforces | DZ_DasherKTB | Active — solving, grinding |
| 🟧 LeetCode | DSA + Competitive | Regular practice |
Appeared in Flipkart GRID 7.0 — survived the brutal CP rounds under time pressure.
"The CP rounds were brutal under time pressure, but a fun challenge nonetheless."
🏛️ IIT Jammu — B.Tech CSE (Hons) | CGPA: 8.87/10 | 2024–2028
📚 Class 12 (CBSE Non-Medical) | 92.0% | 2024
📚 Class 10 (CBSE) | 96.4% | 2022
Positions:
- 🧑💻 Co-Head — Coding Club / AI-ML Community, IIT Jammu
- 🔭 Team Lead — GeoST-Net Research Team (ISRO Hackathon)
[ ] Post-GPU AI Hardware (TOU Framework) — with Dr. Uma Satya Ranjan
[ ] Multimodal architectures for bioinformatics + spatial transcriptomics
[ ] Agentic systems with structured reasoning (Planner-Filler-Validator)
[ ] Instruction-conditioned any-to-any modeling
[ ] VPN infrastructure & network security tooling
[ ] Assembly / low-level systems programming
[ ] GNN-augmented medical imaging
╔══════════════════════════════════════════════════════════╗
║ trying everything. working on stuff. not stopping. ║
║ ║
║ build → break → rebuild better → repeat ║
╚══════════════════════════════════════════════════════════╝
Last updated: 2026 — still going