📊 Stats · 👋 About · 🛠️ Tech Stack · 🚀 Projects · 🔬 Research · 🗺️ Roadmap
const harshit: Developer = {
name: "Harshit Subramanian",
role: "Software Development Engineer (SDE-III) // AI & ML Specialist",
location: "📍 Coimbatore, India",
focus: [
"🧠 Architecting Custom Small Language Models (SLMs) & Hybrid Encodings",
"📚 Building zero-hallucination Retrieval-Augmented Generation (RAG) Systems",
"⚙️ High-precision AST parsing for data engineering workflows",
"⚡ High-performance algorithmic optimization & system design",
],
stack: {
languages: ["Python", "Java", "SQL", "Bash"],
backend: ["Streamlit", "Pandas", "Seaborn"],
ai_ml: ["PyTorch", "Hugging Face", "LangChain", "ChromaDB", "PEFT"],
devOps: ["Docker", "Git", "Linux"],
},
leisure: ["🏋️ Fitness & Body Recomposition", "🎮 Gaming", "🎵 Audio Immersion", "⚡ Competitive Programming"],
principle: "Optimized · Deterministic · Architected for Scale",
};Languages & Frameworks
AI / ML & Tooling
| Project | Description | Tech | Source |
|---|---|---|---|
| 🧠 NOVA-Verse SLM | Custom 14M parameter Small Language Model utilizing a proprietary hybrid positional encoding mechanism. Fine-tuned on Project Gutenberg. |   |
Model |
| 📚 DocMentor RAG | Enterprise-grade Retrieval-Augmented Generation tutor system. Features custom document mentoring with real-time vector embeddings. |   |
Repo |
| ⚙️ SQL Table Extractor | Robust parsing utility leveraging sqlglot. Capable of high-precision AST processing to isolate base tables/columns while ignoring aliases and CTEs. |
  |
Repo |
| 📊 Expense Analytics | Full-stack financial visualization engine processing JSON entries into dynamic time-series filters, stacked bars, and donut charts. |   |
Repo |
| 🌴 TripWise | Travel organization app featuring smart routing, budget estimation, and a custom T-shaped palm tree brand aesthetic. |   |
Repo |
| ⚡ Optimization Suite | Production-grade repository of advanced DSA implementations focused on runtime optimization, graph traversals, and dynamic programming. |   |
Repo |
My core research focuses on pushing the boundaries of parameter efficiency, custom tokenization, and contextual retention in Language Models.
|
📄 NOVA & NOVA-Verse: Designing a 14M Parameter Language Model with Custom Hybrid Positional Encodings Open Research & Model Architecture Documentation · Sep 2025 Engineered a proprietary hybrid positional encoding mechanism for optimized context retention, proving efficiency at a micro-scale. |
📄 Architectural Innovations in Small Language Models (SLMs): Parameter-Efficient Training Technical Report & Open Source Implementation · Oct 2025 Deep-dive lab exploring custom tokenization on the Project Gutenberg corpus and the application of Parameter-Efficient Fine-Tuning (PEFT). |
graph TB
A[🎯 Focus Areas] --> B[🧠 SLM Architectures]
A --> C[⚙️ Agentic RAG Systems]
A --> D[📊 AST Data Engineering]
B --> B1[Scaling Hybrid Positional Encodings]
B --> B2[Parameter Efficient Fine-Tuning]
C --> C1[LangChain / Multi-Agent Orchestration]
C --> C2[ChromaDB Multi-Modal Indexing]
D --> D1[sqlglot Engine Optimization]
D --> D2[Zero-Hallucination Query Parsing]
style A fill:#00F2FE,stroke:#000,color:#000
style B fill:#161b22,stroke:#20BEFF,color:#fff
style C fill:#161b22,stroke:#A8FF60,color:#fff
style D fill:#161b22,stroke:#FFD21E,color:#fff
