Skip to content
View instax-dutta's full-sized avatar
😇
Hi there !!
😇
Hi there !!

Block or report instax-dutta

Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
instax-dutta/README.md

Sai Dutta Abhishek Dash

AI Engineer · On-Device ML · Security · Language AI

AI Engineer from India — building language AI for 38M Odia speakers, on-device ML models for edge security, and privacy-first infrastructure. Also working on graph-native entity investigation and self-verifying AI agents.

Portfolio Blog Email


LinkedIn Twitter Hugging Face


About Me

I'm Sai Dutta Abhishek Dash — an AI Engineer based in Odisha, India.

I build at the intersection of language technology, on-device ML, and privacy-preserving infrastructure. My work spans model optimization (abliteration, distillation, quantization), AI agent tooling, security research, and self-hosted platforms. I maintain 46+ models and 22+ datasets on Hugging Face, and ship open-source projects that power production systems.

All projects, case studies, and services: sdad.pro
Blog: blog.sdad.pro


What I Build

Domain Focus
On-Device ML Sub-100KB inference models, edge deployment, model abliteration and quantization
Security Phishing detection, forensic tooling, network monitoring, scam prevention
Language AI Tokenizer optimization, model distillation, speech and translation for low-resource languages
AI Agents Self-verifying coding agents, agent skills, workflow automation
Data Infrastructure Graph-native entity resolution, provenance-first pipelines, public records indexing
Privacy Zero-knowledge architecture, on-device processing, data sovereignty


Track Record

85+ Repos 46 HF Models 22 HF Datasets AWS Certified
Open source shipped Published on HF Curated & public Cloud Practitioner

Currently Building

const sdad = {
  name:     "Sai Dutta Abhishek Dash",
  alias:    "instax-dutta / R4C3R",
  website:  "https://sdad.pro",
  hf:       "https://huggingface.co/saidutta69",

  currentlyBuilding: [
    "On-device ad blocker model for edge security",
    "AgentLoop — self-verifying autonomy wrapper for coding agents",
    "Odia-exclusive LLM family — tokenizer, model distillation, speech",
  ],

  recentlyShipped: [
    "Packet-Buddy — cross-platform network usage monitor",
    "PhishScout  — 131.5 KB on-device phishing URL detector",
    "Forensic-Recovery — PowerShell forensic evidence acquisition tool",
    "RedPillV1 — 1B parameter drug discovery fine-tune (Hugging Face)",
  ],

  openFor:    ["Technical collaborations", "AI infrastructure consulting", "Interesting projects"],
  contact:    "contact@sdad.pro"
};

Featured Projects

On-Device ML & Edge Inference

Project What it does
PhishScout 131.5 KB on-device phishing URL detector. 90% precision, ~4ms per lookup. Powers network-level scam protection.
AdVig On-device ad blocker model for edge deployment.
RaceBench Making sub-3B models capable enough to replace frontier APIs in edge deployments.
RedPillV1 Drug discovery and formulation fine-tuning pipeline.

Security & Data Infrastructure

Project What it does
PhishTrap Auto-refreshing phishing URL dataset. Fetches OpenPhish/Phishing.Database/PhishStats every 6h. 20k samples, 667 downloads.
Forensic-Recovery Forensically-sound digital evidence acquisition. PowerShell, SHA-256.
Packet-Buddy Cross-platform network usage monitor with real-time analytics.

AI Agents & Infrastructure

Project What it does
AgentLoop Harness-agnostic, self-verifying autonomy wrapper for coding agents. Drives OpenCode/Kilo/Claude/Codex in a verify-gated loop with a held-out oracle.
MarkItDownJS Universal document to Markdown conversion engine (PDF, DOCX, PPTX, XLSX, HTML, CSV, audio, EPUB). LLM-friendly output.
Ansora Self-hostable, serverless blogging platform. Every post is a Markdown file, every save is a git commit. No database.

AI Agent Skills

Project What it does
Google Code Review Google's official code review best practices, packaged as an AI agent skill.
Market Validator Validate SaaS/product ideas by scraping real user complaints across 10+ platforms.
Brand Vibes Apply any company's design language to your website. 66 brand profiles (Stripe, Linear, Vercel, Apple, Claude...).
GSD Skills GSD workflow skills for opencode — plan, execute, ship, code-review, debug.
Only Skills You Need One install command for every coding AI tool.

Tech Arsenal

Languages & Frameworks

Python TypeScript JavaScript C++ C Rust PowerShell Shell React Next.js Node.js Tailwind CSS

AI & ML

PyTorch Hugging Face Ollama TensorFlow scikit-learn Google Vertex

Infrastructure

AWS Docker Git Linux Vercel Netlify Kubernetes


Certifications

Certification Issuer Year
AWS Cloud Practitioner AWS 2024
AWS Bedrock — Build a Gen AI Chatbot Udemy 2024
Amazon Bedrock Getting Started AWS 2024
Amazon Q Developer AWS 2024
Planning a Machine Learning Project AWS 2024
Python for Data Science & Machine Learning Udemy 2024
Project Management LearnTube 2024

GitHub Analytics

GitHub Streak — instax-dutta

Let's Connect

Open to technical collaborations and interesting projects.

Start here: sdad.pro


LinkedIn Twitter Hugging Face GitHub Email Discord


Buy Me A Coffee


AI Engineer · Language AI · On-Device ML · Security sdad.pro · Odisha, India

Profile Views

Pinned Loading

  1. Ultimate-Firewall Ultimate-Firewall Public

    Stop DDoS attacks before they take your server down.

    Shell 11 1

  2. MarkItDownJS MarkItDownJS Public

    Universal document-to-Markdown conversion engine for TypeScript/JavaScript Convert PDFs, DOCX, PPTX, XLSX, HTML, CSV, JSON, XML, images, audio, EPUB, and archives into clean, LLM-friendly Markdown.

    TypeScript 1 1

  3. Auralyn Auralyn Public

    Discord music bot with integrated Lavalink server, ready for deployment on Pelican/Pterodactyl panels.

    JavaScript

  4. binify binify Public

    Binify is a sovereign, production-ready, privacy-first pastebin application. Built for the modern web with a zero-knowledge architecture, ensuring that your secrets remain yours — even from the ser…

    TypeScript

  5. cl-chat-reborn cl-chat-reborn Public

    Peer-to-peer command-line chat with per-connection ECDH encryption, mesh message propagation, and full input sanitization. No servers, no accounts, no logs.

    HTML

  6. packet-buddy packet-buddy Public

    A lightweight, cross-platform network usage monitor that runs silently in the background and provides beautiful real-time analytics.

    Python 2