Third-year Computer Science & Engineering student at Synergy Institute of Engineering and Technology, Dhenkanal, with deep hands-on experience shipping production-grade open-source tools across Rust systems tooling, Python AI agents, Windows/Linux/macOS endpoint security, and offline-first Android apps.
Catching malicious behaviors in user-space is already too late. Sentinel Antivirus and InstallScope operate at the syscall boundary using Linux eBPF, strace, fanotify, FSEvents, ETW, and Windows kernel minifilter concepts to intercept credential reads, volatile memory injections, and unauthorized process spawns.
Autoregressive LLMs are slow and poorly calibrated on fast coding options. Cortex-1 Large trains a non-autoregressive ModernBERT-Large (421M) prefrontal cortex with dynamic marker-token pooling — executing single forward passes in ~32.8 ms on consumer GPU with 79.73% accuracy vs 45.50% API baseline (Brier 0.097 vs 0.369).
Running heavy deep learning on every routine file wastes CPU cycles. Sentinel Antivirus filters >98% of routine files in under 5 ms (Hash Cache → Authenticode → YARA → Lightweight PE), running 2,568-dimensional EMBER2024 LightGBM inference (~180 ms) exclusively on suspicious binaries.
From DBAgent's sqlglot AST + EXPLAIN verification gates and confidence circuit-breakers to Agent-Commerce Trust Auditor and Fake Urgency Checker, every tool is built with read-only defaults, SSRF protection, and formal verification.
All 8 independently designed open-source projects from github.com/mukti-sys — shipped end-to-end from architecture and model training to CI pipelines, PyInstaller/systemd packaging, and published Hugging Face weights.
Intent reconstruction with glossary, embedding schema retrieval + drift detection, sqlglot AST validation, EXPLAIN verification, confidence gate that refuses when uncertain, and concurrency circuit-breaker.
Answers "Is my old phone still healthy?" and "Is anything quietly working against me?" — display/sensor test suites, installed-app permission risk scoring, security-patch EOL warnings, and local trend history.
Fetches product pages via secure SSRF-protected proxy and detects fake countdown timers, phantom "X people viewing" counters, and known dark-pattern plugins (Hurrify, Beeketing, Fomo), returning FAKE / SUSPICIOUS / CLEAN verdicts.
Scans e-commerce storefronts the way an AI shopping agent would — scoring Legibility (schema.org consistency), Consistency (anti-cloaking), and Safety (hidden text & prompt-injection attempts) with a before/after demo store.
Extends Android Digital Wellbeing with up to 2h break intervals, silent background engine, automatic media pause + screen-off + lock on timer expiry via Accessibility Service & exact alarms.
Want to test Sentinel EDR, InstallScope, Cortex-1 & DBAgent live?
EXPLORE GITHUB.COM/MUKTI-SYSSeeking a paid internship (Remote / Odisha / India)? Hire Mukti
Mukti engineers production-grade open-source tools in endpoint security, systems programming, AI agents, and Android development. Built with Rust, Python, PyTorch, and Kotlin, with a strong bias toward measurable benchmarks and least-privilege design. Crafted by github.com/mukti-sys.
Seeking a paid internship (Remote / Odisha / India)? Email mishramukti182@gmail.com or .
3rd Year B.Tech CSE (Expected 2027) • Synergy Institute of Engineering and Technology, Dhenkanal, Odisha
Filters >98% of routine files in <5 ms; runs 2,568-dim EMBER2024 LightGBM (~180 ms) exclusively on suspicious binaries.
Single forward pass with dynamic marker-token pooling: 79.73% SWE-bench/CVE accuracy (Brier 0.097) vs 45.50% API baseline.
Prioritizes honesty over blind execution: Glossary Intent → Schema Drift Check → sqlglot AST → EXPLAIN Plan → Confidence Gate.
Evaluated across 250 real npm packages (840,000+ syscalls). Content-addressed behavioral version-diff eliminates maintainer alert fatigue.
Combines Agent-Commerce Trust Auditor (Legibility, Consistency, Prompt-Injection Safety) with Fake Urgency Checker (SSRF-safe timer detection).