darsh = {
"institute" : "IIT Bhilai",
"degree" : "B.Tech — Data Science & AI",
"focus" : ["LLM Engineering", "Agentic Systems", "Neural Network Interpretability"],
"building" : "Agents that think · Models from scratch · RAG that actually works",
"philosophy" : "Learn by building. Understand by breaking.",
"looking_for": "AI/ML Internship — Agent Systems, LLM Tooling, Applied Research"
}Languages
AI / ML
Backend & Infra
|
Production-ready autonomous agent managing GitHub repos via LangGraph ReAct loop, FastMCP tools, and dual memory (Mem0 + LangGraph).
|
Agentic blog pipeline with real-time web research, citation grounding, and parallel section writing - across 3 progressive versions.
|
|
TypeScript CLI AI agent powered by Groq that interacts with the filesystem using natural language commands.
|
Demand-aware AI API router that automatically switches to a backup provider when traffic exceeds threshold - inspired by Gemini's high-demand behaviour.
|
|
Modular DeepSeek LLM architecture - Byte-level BPE tokenization, Multi-Head Latent Attention (MLA), and Mixture-of-Experts (MoE).
|
Decoder-only GPT Transformer from scratch - multi-head causal self-attention, positional embeddings, autoregressive generation. Zero HuggingFace.
|
|
Clean PyTorch LoRA implementation - understanding how low-rank weight updates enable efficient fine-tuning from first principles.
|
Isolates the pure effect of distillation - teacher-guided student vs hard-label baseline on identical architectures with 5-fold cross-validation.
|
|
Hands-on journey through real software internals - memory allocators, shells, interpreters, emulators, and quantum algorithms. |
|
CRAG pipeline with multi-stage corrective layer - document grading → strip extraction → filtering → context refinement → verified generation.
|
Deep semantic activation analysis for PyTorch/HuggingFace models - probe any layer across 6 classifier types to discover where concepts are encoded.
|
|
Production-style LLM safety pipeline - regex filters, ML toxicity classifiers, prompt injection detection, and output scanning. Inspired by Anthropic, OpenAI & Meta.
|
|
Production-grade LLM inference platform - BullMQ job queue, SSE token streaming, RAG pipeline (Qdrant + Gemini embeddings), and JWT auth.
|
Batteries-included dataset quality testing framework - completeness, validity, distributions, label noise, drift, and text quality with unified scoring reports.
|
|
TypeScript REST API with JWT authentication and role-based authorization using Express, bcrypt, and jsonwebtoken.
|
