┌─────────────────────────────────────────────────────────┐
│ │
│ ▄▀█ █▀█ █▄█ ▄▀█ █▄░█ █▀ █░█ ▄▀█ █▀█ █▀▄▀█ ▄▀█ │
│ █▀█ █▀▄ ░█░ █▀█ █░▀█ ▄█ █▀█ █▀█ █▀▄ █░▀░█ █▀█ │
│ │
│ CS Undergrad · Applied AI & Systems │
│ │
└─────────────────────────────────────────────────────────┘
I build backend systems and applied ML pipelines — retrieval systems, LLM inference workflows, and evaluation-driven architectures. Currently focused on hybrid search and RAG systems. I read papers, implement them, and validate results.
aryan = {
"role" : "Applied AI/ML · Backend · Systems",
"education" : "B.Tech CSE · JIIT Noida",
"building" : ["LLM inference pipelines", "hybrid retrieval systems", "RL environments"],
"open_to" : ["SDE", "Backend Engineering", "AI/ML Engineering", "Data Engineering"],
}| Domain | Technologies |
|---|---|
| Languages | Python · C/C++ · SQL |
| Backend | FastAPI · Async Python · REST · WebSockets · Pydantic · SQLAlchemy · PostgreSQL |
| AI / ML | NLP · Transformers (BERT, FlanT5) · Scikit-learn · Reinforcement Learning · spaCy |
| Retrieval | BM25 · FAISS · Qdrant · Dense + Sparse Search · RRF · CrossEncoder Reranking |
| LLM Tooling | OpenAI SDK · Groq API · LangChain · RAG Pipelines · Prompt Engineering |
| Infra & Dev | Docker · Git · pytest · Linux CLI · SQLite · MySQL |
🔹 InventOps — RL Supply Chain Simulation Engine
Stack: Python · FastAPI · Groq (Llama-3.1) · OpenAI SDK · Docker · pytest · Pydantic v2
OpenEnv-style RL environment for multi-echelon supply chain optimization
- Modeled stochastic demand across 25 SKUs with multi-warehouse setups and configurable difficulty levels
- Reduced LLM observation token size by 87% via optimized state serialization in a Groq/Llama-3.1 pipeline
- Built as a Dockerized FastAPI service with a 17-test pytest suite for validation
🔹 Scipher — Document ETL & NLP Pipeline
Stack: Python · FastAPI · Docling · BERT · FlanT5 · spaCy · SQLAlchemy · SQLite · Docker
Multi-stage pipeline for converting academic PDFs into structured data
- End-to-end flow: parsing → section classification (BERT, 92% accuracy) → NER (spaCy) → summarization (FlanT5) → glossary generation
- Async FastAPI service with WebSocket support, Pydantic-validated I/O, and integrated logging
🔹 HybridIR — Hybrid Retrieval System (ongoing)
Stack: Python · FastAPI · BM25 · Qdrant · sentence-transformers · CrossEncoder · Groq API · Tavily
Hybrid search system combining sparse and dense retrieval with reranking
- BM25 + Qdrant dense retrieval fused via Reciprocal Rank Fusion (RRF) with CrossEncoder reranking
- Evaluated on labeled queries using Precision@K, MRR, nDCG@5; targeting measurable improvement over BM25 baseline
- Web-augmented retrieval via Tavily; side-by-side retrieval comparison UI via Textual
🔹 TokenFlow + BPE Implementation — Subword Tokenization
Stack: Python · NLTK · Regex · FastAPI · Pydantic · JSON · Pickle
Implementation of Sennrich et al. (2016) BPE for Neural Machine Translation
- Built merge-rule learning, subword segmentation, and vocabulary construction from scratch
- Configurable tokenizer with Unicode (NFKC) normalization, regex-based preprocessing, and pair statistics caching
- Compared OOV handling against word-level and character-level baselines
→ Extending HybridIR: query routing + multi-stage reranking + web-augmented retrieval
→ Evaluating LLM pipelines across latency and output quality tradeoffs
→ Studying IR metrics, vector indexing, and inference optimization
# Open to: SDE · Backend Engineering · AI/ML Engineering · Data Engineering
echo "aryan.shr.04@gmail.com"
open "https://linkedin.com/in/arynshrma"