I'm Deepak Krishna, a machine learning engineer. I fine-tune transformers, build RAG systems and ship computer-vision models, then wrap them in APIs and interfaces people can actually use.
1st / 506
teams · IPR hackathon
88.5%
accuracy · Sentira-3D
01 / LEARNING
Fine-tuned transformers
02 / RAG
Local-LLM retrieval
03 / VISION
Real-time computer vision
04 / AGENTS
n8n workflow automation
// about
I'm a B.Tech student in Artificial Intelligence and Machine Learning at Presidency University, Bengaluru, and a former Summer Intern at IIT Jammu, where I worked on generative AI, AI agents and workflow automation.
I like taking models out of notebooks and turning them into products: a DistilBERT sentiment engine with a 3D front end, a local-LLM RAG platform, a real-time sign-language translator. I also write about what I learn while building agent workflows in n8n.
I enjoy hackathons and datathons, which is where I've picked up a first place out of 506 teams and a runner-up finish at an international tech fest.
0.00 / 10
GPA (B.Tech AI & ML)
1 of 0 teams
1st place, IPR hackathon
0
GitHub repositories
// projects
loading repos…
// experience
Jun 2026 – Aug 2026
Indian Institute of Technology Jammu · On-site
Highlights
Innovation and competition record
Expected Aug 2027
Presidency University, Bengaluru
// skills
// contact
Open to work. Whether it's an ML engineering role or an interesting project, send a message and I'll get back to you.
Real-time sentiment analysis engine. DistilBERT fine-tuned on 1.6M tweets reaches 88.5% accuracy (ROC AUC 0.9546), with INT8 quantization for under 10ms inference on NVIDIA GPUs. FastAPI backend, React + Three.js 3D particle front end.
Explore in 3D
AI scheme-matching engine from the SheLeads 2.0 hackathon (Team CodeCorps). A RAG pipeline over local LLMs (Ollama), Qdrant vector search and MongoDB connects Indian women to the financial schemes that fit them.
Explore in 3D
End-to-end pipeline for a heavily imbalanced classification problem (about 1% fatal cases). Benchmarked five models, picked LightGBM for a 0.506 macro F1, and shipped a Streamlit app with live class probabilities.
Explore in 3D
Real-time sign-language-to-text translator. MediaPipe hand landmarks feed SVM and Random Forest classifiers, reaching 99% inference accuracy.
Explore in 3D
AI-powered marine intelligence platform that brings together oceanographic, climate, fisheries and biodiversity data for decision support.
Explore in 3D
Algorithmic swing trading with reinforcement learning: agents that learn entry and exit policies from market data.
Explore in 3D
Alternative credit scoring for people with thin credit files, from the BNP Paribas InnoVerse hackathon. A transparent 0-1000 rule engine over 15 behavioural features, with explanations, a what-if simulator, product matching and an anonymised lender marketplace.
Explore in 3D
Full-stack ordering system for a smash-burger shack. Pickup and delivery, with live kitchen, dispatch, driver and admin screens, a server-priced checkout, an order state machine and PDF invoices. Payments are mocked.
Explore in 3D
AI companion that helps low-income households make safer money decisions. A deterministic metrics engine computes the risk and the AI only explains it. Built for the BWT Hackathon with React, FastAPI, Clerk and MongoDB.
Explore in 3D
A glassmorphism contact page behind a luggage tag's QR code, so a finder can reach the owner in one tap. Animated rain and blob backdrop with a 3D tilt, all in a single dependency-free HTML file.
Explore in 3D
