# Sai Dutta Abhishek Dash — Full Profile > Founder and engineer building AI infrastructure, language technology for underserved markets, and security systems at production scale. Based in Odisha, India. Ships open-source by default. **Last updated:** September 2026 ## Identity Sai Dutta Abhishek Dash (also known as instax-dutta / R4C3R0P) is a founder and engineer from Odisha, India. He builds across the full stack — from C99 inference engines to Next.js web platforms — with a focus on AI infrastructure, language technology for underserved markets, and privacy-first security systems. He has shipped 22 open-source production systems, maintains 85+ public repositories on GitHub, and has published 46+ models and 22+ datasets on Hugging Face. He is AWS Certified (Cloud Practitioner, 2024) and holds a Bachelor's degree in Computer Science from GIET University Gunupur (2025). ## Core Thesis Three conviction-driven bets: 1. **Language AI should exist for every language with 10M+ speakers.** There are 38 million Odia speakers. Zero commercial Odia-specific AI APIs exist. Maelis Research is building the first one — tokenizer optimization, LLMs, speech recognition, and translation. 2. **On-device ML is the future of security.** Not every threat detection model needs a cloud API. PhishScout is a 131KB model that detects phishing URLs with 90% precision in ~4ms, running entirely on-device. 3. **Privacy-first doesn't mean worse UX.** Binify, CL-Chat Reborn, and CL-Chat prove you can build zero-knowledge systems that feel native. Client-side encryption, no accounts, no telemetry. ## Detailed Product Profiles ### AgentLoop **Category:** AI Agent Infrastructure **Language:** Python **URL:** https://github.com/instax-dutta/agentloop Harness-agnostic, self-verifying autonomy wrapper for coding agents. AgentLoop drives OpenCode, Kilo, Claude, Aider, and Codex in a verify-gated loop with a held-out oracle. The agent must prove its changes are correct before the loop terminates. This solves the fundamental reliability problem in autonomous coding agents — they stop when they're right, not when they give up. ### ornith-flight **Category:** LLM Inference **Language:** C99 **URL:** https://github.com/instax-dutta/ornith-flight Expert-streaming C99 inference engine for Ornith 35B MoE. Runs 20 GB models on 8 GB Apple Silicon via per-layer streaming I/O. Zero Python dependencies, pure systems engineering. Demonstrates that high-performance inference doesn't require GPU clusters — a well-engineered CPU implementation can serve real workloads on consumer hardware. ### MarkItDownJS **Category:** Document Processing **Language:** TypeScript **URL:** https://github.com/instax-dutta/MarkItDownJS AST-first document ingestion and conversion platform. Converts PDFs, DOCX, PPTX, XLSX, HTML, CSV, JSON, XML, images, audio, EPUB, and archives into clean, LLM-friendly Markdown. Built as a Turborepo monorepo with Vite, pnpm, and RAG-ready chunking. Used as the backend engine for Tooner and other AI pipeline tools. ### Vulscany **Category:** Security Engineering **Language:** TypeScript / Next.js **URL:** https://github.com/instax-dutta/vulscany Private, self-hostable AI-powered code security agent. Born from a venture-backed startup codebase, now open-source. Multi-language scanning, AI-powered fix generation, real-time CVE matching, GitHub OAuth integration. Zero database dependencies — runs entirely locally. ### PhishScout **Category:** On-Device ML **Language:** Python **URL:** https://huggingface.co/saidutta69/PhishScout **Repository:** https://github.com/instax-dutta/PhishScout 131.5 KB on-device phishing URL detector. 90% precision, ~4ms per lookup. Trained on curated phishing datasets from OpenPhish, Phishing.Database, and PhishStats. Designed for network-level scam protection — runs on edge devices, home routers, and lightweight VPS instances without cloud dependencies. ### Hinglish-Bench **Category:** NLP Benchmarking **Language:** Python **URL:** https://github.com/instax-dutta/hinglish-bench Reference-free benchmark for measuring LLM Hinglish (Hindi-English code-mixed) text generation. 34 prompts across 8 categories. Metrics include CMI (Code-Mixing Index), M-Index, SyMCoM, plus LLM judge evaluation. Integrates with lm-evaluation-harness for standardized evaluation pipelines. ### PraharShield **Category:** Security Engineering **Language:** Java **URL:** https://github.com/instax-dutta/PraharShield High-performance bot filtering plugin for Minecraft Velocity proxies. Built on LimboAPI. Features CAPTCHA handshakes, client fingerprinting, and pattern-based detection to block 100k join-flood JPS attacks on a 1GB RAM VPS. ### CL-Chat Reborn **Category:** Cryptography / P2P **Language:** Python **URL:** https://github.com/instax-dutta/cl-chat-reborn Peer-to-peer command-line chat with per-connection X25519 ECDH key exchange, ChaCha20-Poly1305 AEAD encryption, and HKDF-SHA256 key derivation. Serverless mesh propagation, TOFU fingerprint verification, full input sanitization. 81 pytest tests covering all cryptographic operations. ### Forensic-Recovery **Category:** Digital Forensics **Language:** PowerShell **URL:** https://github.com/instax-dutta/Forensic-Recovery Forensically-sound digital evidence acquisition tool for Windows. SHA-256 chain-of-custody verification, automated system file exclusion, audit logs, dry-run mode, multi-drive triage. Designed for law enforcement and incident response teams. ### DadGuard **Category:** Windows Security **Language:** C# **URL:** https://github.com/instax-dutta/DadGuard Lightweight Windows security watchdog for protecting non-technical family members. Real-time threat monitoring, suspicious process detection, and alert system — all with minimal system overhead. ### Binify **Category:** Privacy Infrastructure **Language:** TypeScript / Next.js **URL:** https://bin.sdad.pro **Repository:** https://github.com/instax-dutta/binify Zero-knowledge encrypted pastebin with client-side Web Crypto API encryption. Data is encrypted in the browser before transmission. The server never sees plaintext. Built with Next.js, Turso database, and Upstash Redis for rate limiting. ### Ansora **Category:** Developer Platform **Language:** TypeScript **URL:** https://blog.sdad.pro **Repository:** https://github.com/instax-dutta/ansora Self-hostable, serverless-deployable blogging platform. Every post is a Markdown file, every save is a git commit. No database, no SaaS CMS. Deploy to Netlify, Vercel, Docker, or any VPS. Git-native version control for all content. ### The Watcher **Category:** Data Engineering / AI **Language:** Python **URL:** https://github.com/instax-dutta/The-Watcher Reddit used-tech intelligence bot. Monitors 13 Indian used-tech subreddits 24/7, identifies WTS (Want to Sell) posts, uses LLM extraction to parse price, specs, and condition from natural language, and drops rich embeds in Discord channels. ### Visitor Analytics **Category:** Privacy / Developer Tooling **Language:** TypeScript **URL:** https://github.com/instax-dutta/visitor-analytics Production-ready, framework-agnostic analytics SDK. Tracks visitors, page views, performance metrics, and user interactions — all anonymous, zero PII. Privacy-preserving by design. ### PacketBuddy **Category:** Network Monitoring **Language:** Python / JavaScript **URL:** https://github.com/instax-dutta/packet-buddy Cross-platform network usage monitor with real-time analytics. Runs silently in the background, tracks bandwidth usage across interfaces, and provides beautiful real-time dashboards. Zero configuration required. ### Auralyn **Category:** Discord Bot **Language:** JavaScript **URL:** https://auralyn.sdad.pro **Repository:** https://github.com/instax-dutta/Auralyn Zero-infrastructure Discord music bot with a bundled Lavalink server. High-fidelity audio playback, smart search, queue management, and single-container Docker deployment. No external Lavalink setup required. ### 3D Web Contents **Category:** Frontend / Creative Coding **Language:** JavaScript **URL:** https://github.com/instax-dutta/3d-web-contents Zero-dependency, drop-in 3D components for React + Three.js. Eight copy-paste templates: CSS3DCard, ParticleField, MorphingBlob, WarpSpeed, TunnelScroll, GeometryRepel, LightingReflection, BarChart3D. Own the source, skip the npm bloat. ## AI Agent Skills ### Brand Vibes **URL:** https://github.com/instax-dutta/brand-vibes Ships a library of 66 full brand design profiles (Stripe, Linear, Vercel, Apple, Claude, Nike, Ferrari, and more) plus an off-library DNA-extraction workflow. Apply any company's design language to your website while vibecoding. Resolves the brand, extracts design DNA into semantic tokens, applies to Tailwind/CSS/shadcn. ### Market Validator **URL:** https://github.com/instax-dutta/market-validator Validate SaaS and product ideas by scraping real user complaints from Reddit, Twitter, G2, Quora, HackerNews, ProductHunt, V2EX, XiaoHongShu, Bilibili, and YouTube. Uses agent-reach for multi-platform data collection. Replaces guesswork with evidence from actual user pain points. ### Google Code Review **URL:** https://github.com/instax-dutta/google-code-review Google's official code review best practices packaged as an AI agent skill. Covers both reviewer and developer perspectives — what to look for, how to comment, CL descriptions, handling feedback, and resolving conflicts. ### Only Skills You Need **URL:** https://github.com/instax-dutta/only-skills-you-need One install command for every coding AI tool. Curated skills stack: full-stack dev, frontend/React, SEO, workflow automation, and code review. Works on opencode, Claude Code, Antigravity, Kilocode. ### Roadmap Tutor **URL:** https://github.com/instax-dutta/roadmap-tutor Hermes Agent skill: learn any roadmap.sh roadmap one topic at a time, tracked across sessions. Stdlib-only Python, runs on a Raspberry Pi. ## Maelis Research Maelis Research builds language AI infrastructure for Odia and low-resource Indian languages. Founded in 2026, based in Dhenkanal, Odisha. **The opportunity:** 38 million Odia speakers exist. Norwegian (5M), Dutch (25M), and Greek (13M) all have robust AI support. Odia has none. No commercial Odia-specific AI API exists today. **What Maelis is building:** - Odia-optimized tokenizer — 3x more efficient than generic alternatives for Brahmic scripts - Lekhani (ଲେଖନୀ) model family — Odia-exclusive LLMs in capability tiers (Pada, Chhanda, Kavya, Mahakavya) - Shruti (ଶ୍ରୁତି) — Odia speech recognition and synthesis - Anuvada (ଅନୁବାଦ) — Odia translation with cultural context - odia-eval-benchmark — 121,947 rows, 7 task families, 34 datasets (published on Hugging Face) All research released under Apache 2.0. DPIIT Startup Recognition. Positioned for IndiaAI Mission. **Website:** https://maelis.sdad.pro ## Offsage Offsage is a technical agency eliminating operational friction for startups and scaling teams. From first automation to full-scale SaaS. Built in India, deployed globally. Discipline-first web engineering: Next.js, React Server Components, mobile-first responsive design, conversion-optimized layouts, measured animation systems. **Website:** https://offsage.com ## Contact - **Email:** contact@sdad.pro - **GitHub:** https://github.com/instax-dutta - **LinkedIn:** https://www.linkedin.com/in/sdabhishekdash/ - **X (Twitter):** https://x.com/abhishekdash69 - **Portfolio:** https://sdad.pro - **Blog:** https://blog.sdad.pro - **Hugging Face:** https://huggingface.co/saidutta69 Open to: Investment conversations, strategic partnerships, technical collaborations, advisory roles.