B.Tech in Data Science, UPES Dehradun (Class of 2026)
Specialized in Multimodal Deep Learning, Temporal Graph Neural Networks (GNNs), Inductive Conformal Prediction, Causal Inference, and High-Throughput Streaming MLOps.
I engineer mathematically bounded, production-scale AI systems designed to operate under concept drift and non-stationary real-time workloads.
Enterprise-Scale Multimodal Outrage Forecasting, Topological Threat Intelligence & Proactive Defense
Technologies: PyTorch, PyG (GATv2), Mistral-7B, Qwen2-VL, Apache Flink CEP, Apache Kafka, Neo4j GDS, Triton, Conformal Prediction
- Multimodal AI Engine: Combines Mistral-7B-Instruct (4096-dim emotion & moral conviction scoring) with Qwen2-VL NaViT (3584-dim sarcastic meme vision-language parsing) via gated cross-attention fusion.
- Topological Cascade Forecasting: Uses PyTorch Geometric GATv2 over dynamic Neo4j interaction graphs to compute edge-weighted virality trajectories up to 6 hours into the future.
- Statistical Error Guarantees: Implements Inductive Conformal Prediction guaranteeing distribution-free finite-sample error bounds ( - οΏ½lpha = 0.90$) under severe concept drift.
- Sub-50ms Streaming: Processes live cross-platform firehoses through Kafka and Flink Complex Event Processing (CEP) sliding windows dispatched to NVIDIA Triton Inference Server.
- Causal Discovery & Red-Teaming: Models Pearl's Structural Causal Models for platform intervention discovery and stress-tests model robustness against AutoDAN / GCG gradient token perturbations.
- π Documentation & Live Architecture: killer1panda.github.io/prsi
Autonomous Zero-Touch Cloud Disaster Recovery with $0.00 Standing Compute
Technologies: Python, Terraform, Ansible, AWS (Route 53, SSM, S3, EC2), PostgreSQL WAL Streaming
- Autonomously detects on-prem outages via multi-signal quorum, spins up ephemeral AWS replica infrastructure with Terraform & Ansible, replays PostgreSQL WAL streams, cuts over Route 53 DNS, and executes zero-data-loss reverse synchronization on primary recovery.
- Measured telemetry: RPO: 290s (4.8 min), RTO: < 7 min with pre-baked AMIs, $0.00/month standing compute cost.
Low-Latency Conversational AI for Financial Services (Hindi / English / Hinglish)
Technologies: Python, FastAPI, Groq Whisper, Sarvam AI, pgvector RAG, NVIDIA TTS, Neural VAD
- Real-time spoken dialogue engine with dual-gated barge-in, sub-second latency, pgvector semantic search, and dynamic routing across Groq 8B/70B and Sarvam streaming TTS.
π BasketIQ
Quick-Commerce Multi-Cart Optimization Engine
Technologies: Python, FastAPI, PostgreSQL, Flutter, Nemotron-3-nano-omni, Local-First Architecture
- Multi-platform grocery basket solver comparing Blinkit, Zepto, and Instamart feeds with LLM intent parsing and dynamic price-splitting algorithms.
Autonomous Multi-Agent Development Pipeline & Specialized Agent Swarms
Active contributions to autonomous developer toolchains, multi-agent state machines, and system orchestrators.
| Domain | Technologies & Frameworks |
|---|---|
| Data Science & ML | PyTorch, PyTorch Geometric (PyG), Hugging Face Transformers, Scikit-Learn, NumPy, Pandas, DVC |
| Generative AI & LLMs | Mistral-7B, Qwen2-VL, LoRA/QLoRA Fine-tuning, TensorRT-LLM, Triton Inference Server, LangChain |
| Statistical Modeling | Inductive Conformal Prediction, Causal Inference (SCM), Concept Drift Detection (Wasserstein/K-S) |
| Streaming & Data Infra | Apache Kafka, Apache Flink CEP, Neo4j Graph Data Science (GDS), Redis, PostgreSQL, pgvector |
| Systems & MLOps | Docker, Kubernetes (AWS EKS), Terraform, Ansible, GitHub Actions CI/CD, Linux, HPC (PBS Scripts) |
| Languages | Python (Advanced), SQL, TypeScript, Bash/Shell, Go (Fundamentals) |

