Nishanth S

Internships

  • AI Intern

    BOCK

    Bengaluru · Hybrid

    Built a document-intelligence system that turns healthcare documents into structured, usable data.

    April - July 2026

    Python, PyTorch, Hugging Face, Vision-Language Models, LoRA, 4-bit Quantization, Gradient Checkpointing

    • Extracted page-level data from medical reports, clinical records and textbooks with a 4-bit quantized vision-language model.
    • Fine-tuned an 8B LLM for the healthcare domain using LoRA and 4-bit quantization.
    • Designed a two-tier inference setup - a fine-tuned local model first, falling back to larger vision-language models on GPU or CPU - so extraction keeps running on limited hardware instead of failing.
    • Made multi-hour batch extraction jobs resumable, so an interrupted run picks up exactly where it stopped instead of reprocessing the whole corpus.
    • Shipped a fully offline inference client that loads the base model and fine-tuned adapter from disk, with configurable sampling and no network dependency.
  • Internship Training

    Hindustan Aeronautics Limited

    Bengaluru · On-site

    Built a desktop system for managing and analyzing aircraft LRU test data.

    January - April 2026

    Python, PySide6, SQL, MVC, On-device LLM

    • Developed the application in Python (PySide6, SQL) on a Model-View-Controller architecture, keeping the data layer, business logic and interface independent so each could change without breaking the others.
    • Built real-time KPI dashboards with multi-column filtering over the complete test history, replacing manual spreadsheet lookups of aircraft LRU test records with instant cross-referencing.
    • Integrated an on-device LLM assistant that answers plain-English questions from the database's own records, so engineers can analyze test data without writing SQL and get answers grounded in the actual data.
    • Automated data ingestion from source to store, with input validation and error handling at every entry point.

Projects

Click a card to flip it

Live in Production

Forecasting agent

Agentic Finance System

Python / FastAPI / Uvicorn

Agentic Finance System

An AI finance assistant that forecasts crypto and stock moves with honest confidence ranges, lets an LLM fact-check every call against the news and tracks its own record in public. Paper trading included.

Python · FastAPI · Uvicorn · APScheduler

Multi-Agent LLM

Analytics agent

Multi-Agent NL to SQL

Python / DuckDB / sqlglot

Multi-Agent NL to SQL

A five-agent pipeline that turns plain-English questions into DuckDB SQL, with AST validation, grain-aware fan-out checks, row-level security, PII masking, a hash-chained audit log and a 137-case eval suite gated in CI.

Python · DuckDB · sqlglot · FastAPI

Production RAG

RAG system

RAG Document Assistant

Python / FastAPI / React

RAG Document Assistant

Production-grade RAG for document Q&A: cited answers, abstention when evidence is weak and calibrated confidence, over hybrid pgvector and full-text search with cross-encoder reranking.

Python · FastAPI · React · PostgreSQL

19 MCP Servers

MCP framework

Framewright

Python / MCP / FFmpeg

Framewright

An AI-assisted video editing framework built on MCP: 19 Python servers expose 45 tools for media analysis and preparation, so an AI agent can build frame-accurate DaVinci Resolve timelines.

Python · MCP · FFmpeg · DaVinci Resolve

Repo ↗
Database Internals

In-memory database

In-Memory SQL Engine

C# / .NET / BenchmarkDotNet

In-Memory SQL Engine

An in-memory SQL query engine written from scratch in C#, with a B+ tree and hash index, a cost-based query planner and Volcano-style execution with parallel aggregation. Results are verified against SQLite on 500 randomly generated queries.

C# · .NET · BenchmarkDotNet

Live 3D Site

Landing page

Interactive 3D Web Experience

JavaScript / three.js / GSAP

Interactive 3D Web Experience

An interactive 3D landing page with a WebGL wireframe house, scroll-driven storytelling and a page-turning sketchbook.

JavaScript · three.js · GSAP · Lenis

Let's-ConnectLet's-ConnectLet's-ConnectLet's-ConnectLet's-ConnectLet's-Connect