About Utkarsh Jaiswal

Hi, I'm Utkarsh aka Zyaxxy a full-stack blockchain developer building at the intersection of Web3, distributed systems, and full-stack development, with a curiosity for security. I'm a Turbin3 graduate and a Tata HackQuest Season 10 CTF finalist.
I'm currently building Kite, an investing platform on Solana, and exploring projects in DeFi, on-chain protocols, and AI. I enjoy turning complex ideas into useful, intuitive experiences.

Tools I think in

Languages
RustTypeScriptSolidityPythonGo
Interfaces
ReactNext.jsReact NativeTailwind
Systems
SolanaAnchorNode.jsExpress
Data & tools
PostgreSQLConvexDockerAWS
Utkarsh's GitHub profile image
THE OTHER SIDE OF THE BRAIN

Side Projects(12)

FIG. 001 — AN EXPLORATION
+ HOVER TO DECODE
FORM FOLLOWS CURIOSITY
01
12
0112

SOLANA / FANTASY SPORTS

Athlete tokens, live contests, and settlement on-chain.

SCROLL TO EXPLORE
A CONTINUOUS EXPERIMENT.
Back to about
CV Book a call
05 / 12AI / RETRIEVAL & EVALUATION

Voice RAG

A closer look at every answer.

A voice RAG benchmark and dashboard for Hindi, English, and Bengali, bringing streaming retrieval, chunking experiments, and latency evaluation into one workspace.

FIG. 005 / VOICE RAGA closer look at every answer.

Overview.

Built around the MSMARCO-XI dataset, Voice RAG makes retrieval experiments visible. A Next.js dashboard connects dataset exploration and chunking comparisons with a Rust streaming backend and Python ingestion and evaluation pipelines.

From speech to supporting context.

An Axum WebSocket server connects speech-to-text, vector retrieval, and language-model generation. The pipeline supports Hindi, English, and Bengali, with Qdrant integration and a local passage-store option.

Measure the tradeoffs.

Fixed-window, sentence-boundary, metadata-aware, and hierarchical chunking strategies can be compared in one interface. The evaluation tooling tracks retrieval quality and latency percentiles, with a 200 ms target used as an experiment goal.

The system at a glance.

  1. 01Stream speech
  2. 02Retrieve passages
  3. 03Generate an answer
  4. 04Evaluate the run
Keep exploringFaceID Onchain