Projects

Built for systems reliability, low-level performance, and applied ML. Full source at github.com/snigenigmatic.

qwen_cuda

From-scratch CUDA C++ inference engine for Qwen3-8B, no frameworks. GGUF parsing, custom kernels, GQA with KV cache, flash attention, and CUDA graph capture in ~1600 lines. ~37 tok/s on an RTX 4090 at ~17.1 GB peak VRAM.

CUDA · ML · Optimization · Performance

UncertaRAG

Epistemic uncertainty-aware RAG on Mistral-7B. 12-experiment grid reaching AUROC 0.815 under INT4 quantization.

Python · RAG · Uncertainty Quantification · Research

Rakshak-AI

Multi-agent compliance system on LangGraph with a knowledge graph core. 1st runner-up at GDG Cloud Bengaluru Deployfest 2026.

Python · LangGraph · Multi-Agent · Knowledge Graphs

MELA

Offline system for reconnecting missing persons with their families using on-device Gemma. Top 6 finalist at the Google DeepMind Bangalore Hackathon, built in one day with Team Ring Zero.

Python · On-Device ML · Gemma · Hackathon Finalist

LocalLens

MCP-native multimodal file search built on jina-embeddings-v5-omni, LanceDB, and FastAPI, with a Linux/WSL port.

Python · MCP · Embeddings · LanceDB

aloo

Rust-based safety auditing layer for npm/npx packages. Two-tier architecture: a deterministic analysis pipeline plus an LLM intelligence layer.

Rust · Supply Chain Security · LLM · Tooling

cluster-simulator

Implemented Kubernetes-style control-plane logic in safe Rust with scheduling, health checks, and leader-follower coordination.

Rust · Distributed Systems · Control Plane · CI

redis-from-scratch

Built RESP3 compatibility with Pub-Sub and used it as a backing store for a REST API based Kafka clone.

Go · RESP3 · In-Memory Store · Systems

sql-engine

An educational SQL database engine built from scratch in C++ to understand database internals.

C++ · Database Systems · Query Processing · Optimization

orderbook

Benchmarked matching engines to nanosecond-level latencies across x64 and ARM using C++20 and custom timing utilities.

C++20 · High Performance · Low Latency · Benchmarking

YAK

Designed a Kafka-style event stream with replication, offsets, and leader-follower architecture.

Python · Distributed Systems · Pub-Sub · Replication

code-ronin

Won Agentathon '26 by turning debugging practice into an adversarial AI game loop.

Python · AI Agents · Developer Education · Gamification

MediChain-FL

Reached 93% pneumonia detection accuracy with federated training and no raw patient data exposure.

Python · PyTorch · Flower · CKKS Encryption · Ethereum

InsightInvest

Automated multi-source company analysis and generated structured investment outlooks through an agentic workflow.

Python · LLM Agents · Finance Research · Structured Reports