Projects

Shipped Products

Production systems spanning AI infrastructure, language technology, security, developer tooling, and AI agent skills. Built with conviction, shipped at scale. Explore the engineering toolkit behind these, or discuss collaboration.

ornith-flight

C99 Inference Engine for 35B MoE on Consumer Hardware

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, pure systems engineering.

C99MoEApple SiliconInferenceSystems Programming
Code

AgentLoop

Self-Verifying Autonomy Wrapper for Coding Agents

Harness-agnostic, self-verifying autonomy wrapper for coding agents. Drives OpenCode/Kilo/Claude/Aider/Codex in a verify-gated loop with a held-out oracle — so the agent actually finishes and proves it is correct.

PythonLLM AgentsVerificationAutonomy
Code

PhishScout

131 KB On-Device Phishing URL Detector

90% precision, ~4ms per lookup. Sub-100KB on-device ML model for network-level scam protection. Trained and benchmarked with reproducible pipeline on Hugging Face.

PythonOn-Device MLModel QuantizationEdge Inference
ViewCode

Vulscany

AI-Powered Code Security Agent

Private, self-hostable AI code scanner with automated fix generation. Multi-language scanning, real-time CVE matching, and GitHub OAuth with zero database dependencies.

Next.jsTypeScriptMistral AIOllamaGitHub OAuth
Code

CL-Chat Reborn

P2P Encrypted Command-Line Chat

Peer-to-peer command-line chat with X25519 ECDH encryption and ChaCha20-Poly1305 AEAD. Serverless mesh propagation, TOFU fingerprint verification, full input sanitization.

PythonX25519 ECDHChaCha20HKDF-SHA256Cryptography
Code

MarkItDownJS

Universal Document-to-Markdown Conversion Engine

AST-first document ingestion and conversion platform for TypeScript. Converts PDFs, Office docs, images, audio, and archives into structured AI-ready Markdown with RAG chunking.

TypeScriptNode.jsTurborepopnpmVite
Code

Binify

Zero-Knowledge Encrypted Pastebin

Zero-knowledge encrypted pastebin with client-side encryption ensuring absolute privacy. Built with Next.js, Turso, Upstash Redis, and Web Crypto API.

Next.jsTursoUpstash RedisWeb Crypto APITypeScript
ViewCode

Hinglish-Bench

Reference-Free Benchmark for Hinglish LLM Generation

34 prompts, 8 categories, CMI/M-Index/SyMCoM metrics with LLM judge. lm-evaluation-harness integration for evaluating Hindi-English code-mixed text generation.

PythonNLPBenchmarkingLanguage Evaluationlm-eval-harness
Code

Forensic-Recovery

Forensically-Sound Digital Evidence Acquisition

PowerShell digital evidence acquisition tool with SHA-256 chain-of-custody verification. Automated system file exclusion, audit logs, dry-run mode, and multi-drive triage.

PowerShellSHA-256Digital ForensicsWindows
Code

PraharShield

Minecraft Velocity Anti-Bot Filter

High-performance bot filtering plugin for Minecraft Velocity proxies. CAPTCHA handshakes and client fingerprinting to block 100k join-flood JPS attacks on 1GB RAM VPS.

JavaVelocityLimboAPIGradle
Code

The Watcher

Reddit Used-Tech Intelligence Bot

Stalks 13 Indian used-tech subreddits 24/7, snipes WTS posts, LLM-extracts price/specs/condition, and drops rich embeds in Discord. Automated market intelligence.

PythonReddit APILLM ExtractionDiscordAutomation
Code

DadGuard

Windows Security Watchdog for Family Members

Lightweight Windows security watchdog for protecting non-technical family members. Real-time threat monitoring with minimal system overhead.

C#Windows SecurityReal-time Monitoring
Code

Ansora

Self-Hostable Serverless 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 your own VPS.

TypeScriptGitServerlessMarkdown
ViewCode

Visitor Analytics

Privacy-Preserving Analytics SDK

Production-ready, framework-agnostic analytics SDK for web applications. Tracks visitors, page views, performance, and interactions — all anonymous, zero PII.

TypeScriptAnalyticsPrivacy-FirstFramework-Agnostic
Code

Auralyn

Discord Music Bot with Integrated Lavalink

Zero-infrastructure Discord music bot with a bundled Lavalink server. High-fidelity playback, smart search, and single-container Docker deployment.

JavaScriptDiscord.jsLavalinkDocker
ViewCode

3D Web Contents

Zero-Dep React Three.js Components

8 copy-paste 3D components for React + Three.js — CSS3DCard, ParticleField, MorphingBlob, WarpSpeed, TunnelScroll, GeometryRepel, LightingReflection, BarChart3D.

ReactThree.jsWebGL3D Graphics
Code

PacketBuddy

Cross-Platform Network Usage Monitor

Ultra-lightweight bandwidth monitor and network traffic tracker with real-time analytics. Runs silently in the background with zero configuration.

PythonJavaScriptShellNeonDBChart.js
Code

Brand Vibes

Apply Any Company's Design Language While Vibecoding

AI agent skill: ships a library of 66 full brand design profiles (Stripe, Linear, Vercel, Apple, Claude, Nike, Ferrari...) plus an off-library DNA-extraction workflow.

Design SystemsBrand DNATailwind CSSAI Skills
Code

Market Validator

Validate Ideas with Real User Complaints

AI agent skill: scrapes real user complaints from Reddit, Twitter, G2, Quora, HackerNews, ProductHunt, V2EX, XiaoHongShu, Bilibili, and YouTube using agent-reach.

Market ResearchData ScrapingAI SkillsMulti-Platform
Code

Google Code Review

Google's Code Review Best Practices as an AI Skill

Google's official code review best practices, packaged as an AI agent skill. Covers reviewer perspective, developer perspective, CL descriptions, handling feedback, and conflict resolution.

Code ReviewAI SkillsEngineering Practices
Code

Only Skills You Need

One Install Command for Every Coding AI Tool

Curated skills stack for daily use: full-stack dev, frontend/React, SEO, GSD workflow, and code review. Works on opencode, Claude Code, Antigravity, Kilocode.

AI SkillsFull-StackMulti-Tool
Code

Roadmap Tutor

Learn Any roadmap.sh Roadline One Topic at a Time

Hermes Agent skill: learn any roadmap.sh roadmap one topic at a time, tracked across sessions. Stdlib-only Python, runs on a Raspberry Pi.

PythonAI SkillsLearning SystemRaspberry Pi
Code

How many products has Sai Dutta Abhishek Dash shipped?

Sai Dutta Abhishek Dash has shipped 22 open-source production systems spanning AI infrastructure, language technology, security, developer tooling, and AI agent skills. The portfolio contains 5 AI and language infrastructure projects, 5 security and systems projects, 6 platform and product projects, 1 creative coding project, and 5 AI agent skills. Flagship systems include ornith-flight, AgentLoop, MarkItDownJS, PhishScout, and Hinglish-Bench in infrastructure, plus Vulscany, PraharShield, CL-Chat Reborn, Forensic-Recovery, and DadGuard in security. Notable results include a C99 inference engine running 20 GB mixture-of-experts models on 8 GB of Apple Silicon memory, and a 131 KB phishing detector with 90% precision at roughly 4 ms per lookup. He maintains 85+ public repositories on GitHub and has published 46+ models and 22+ datasets on Hugging Face. Every project is documented with its language, architecture, and benchmark results.

What is AgentLoop and how does it work?

AgentLoop is a harness-agnostic, self-verifying autonomy wrapper for coding agents, written in Python and published at https://github.com/instax-dutta/agentloop. AgentLoop wraps an existing coding agent and drives it inside a verify-gated loop backed by a held-out oracle. Instead of letting the agent decide it is finished, AgentLoop runs a verification step against an oracle the agent cannot see, and only terminates the loop once that independent check passes. AgentLoop is harness-agnostic: the same wrapper drives OpenCode, Kilo, Claude, Aider, and Codex without modification. The project targets the core reliability failure in autonomous coding agents, which stop when they give up rather than when they are right, producing unverified changes. AgentLoop is written in Python and is designed for teams that cannot afford unreviewed autonomous changes. It is one of 22 open-source production systems shipped by Sai Dutta Abhishek Dash, who maintains 85+ public repositories on GitHub.

What is ornith-flight and what problem does it solve?

ornith-flight is an expert-streaming C99 inference engine for the Ornith 35B mixture-of-experts model, written with zero Python dependencies. ornith-flight runs 20 GB models on 8 GB of Apple Silicon memory through per-layer streaming I/O, so the model never has to be fully resident in RAM. The engine targets mixture-of-experts architectures and streams experts per layer rather than loading the full parameter set, which keeps peak memory within the 8 GB budget. The project demonstrates that high-performance large language model inference does not require a GPU cluster, and that a well-engineered CPU implementation can serve real workloads on consumer hardware. The repository is public at https://github.com/instax-dutta/ornith-flight. Because it streams experts per layer, peak memory stays within the 8 GB budget regardless of the 20 GB model size, and inference needs no GPU cluster. It is written in C99 as pure systems engineering.

What is PhishScout and how accurate is it?

PhishScout is a 131 KB on-device phishing URL detector that achieves 90% precision at roughly 4 milliseconds per lookup. PhishScout is a sub-100KB machine learning model trained and benchmarked with a reproducible pipeline published on Hugging Face at https://huggingface.co/saidutta69/PhishScout, with source code at https://github.com/instax-dutta/PhishScout. Because the model runs entirely on-device, it is designed for network-level scam protection on edge devices, home routers, and lightweight VPS instances with no cloud dependency. The model is trained on curated phishing datasets and is small enough to ship inside a router, a browser extension, or a lightweight server. PhishScout is one of 22 open-source production systems shipped by Sai Dutta Abhishek Dash. PhishScout is one of 22 open-source production systems shipped by Sai Dutta Abhishek Dash, a founder and engineer based in Odisha, India who maintains 85+ public repositories on GitHub. The model is published with a reproducible training pipeline.

What is Hinglish-Bench and what does it measure?

Hinglish-Bench is a reference-free benchmark for measuring Hinglish text generation, where Hinglish is Hindi-English code-mixed text. It is written in Python and evaluates models without requiring a gold-standard reference answer for every prompt. The benchmark contains 34 prompts spread across 8 categories and scores generations using CMI, a Code-Mixing Index, plus M-Index and SyMCoM metrics alongside an LLM judge. Hinglish-Bench integrates with lm-evaluation-harness so it can be dropped into standardized evaluation pipelines. It is one of 22 open-source production systems shipped by Sai Dutta Abhishek Dash, and the repository is public at https://github.com/instax-dutta/hinglish-bench. The project exists because code-mixed generation is common in everyday Indian English usage but poorly represented in benchmarks built only for standard English. Public Hinglish text is also the working language of everyday commerce and support in India, so a benchmark built only on standard English mismeasures how real users actually write.

Which of Sai Dutta Abhishek Dash's projects have a live web app?

Four of the 22 systems shipped by Sai Dutta Abhishek Dash run as live web applications. Binify is a zero-knowledge encrypted pastebin at bin.sdad.pro, where content is encrypted in the browser with the Web Crypto API before transmission so the server never sees plaintext. Ansora is a self-hostable serverless blogging platform at blog.sdad.pro, where every post is a Markdown file and every save is a git commit. Auralyn is a zero-infrastructure Discord music bot at auralyn.sdad.pro, shipping a bundled Lavalink server in a single container. PhishScout is published as a downloadable model at huggingface.co/saidutta69/PhishScout, where its 131 KB weights and reproducible training pipeline are available. The remaining 18 systems are published as source code on GitHub under permissive licences. Visitor Analytics, a privacy-preserving framework-agnostic SDK with zero PII, is published as an installable package, and The Watcher runs continuously against 13 subreddits posting results to Discord.