“Architecting, building, and deploying production-grade AI systems on zero-cost infrastructure. Frugal engineering isn't a constraint — it's a design philosophy.”

I am an independent systems builder specializing in autonomous agent orchestration, zero-cost infrastructure design, and hardware-software convergence. I build end-to-end applications (Python → Node.js → Flutter → Chrome Extensions) that run in production with high reliability and zero or minimal operating cost. No massive budgets, just clean architecture decisions.

Core Technical Arsenal:

• Autonomous Agent Orchestration: Multi-provider fallback networks (3 providers × 18 models × 36 retry paths) with circuit breakers and anti-loop safety.
• Intelligent Browser Automation: Context-aware sniffers, MutationObservers, and document analyzers running at the extension level.
• Edge AI & On-Device Runtimes: Flutter-based local LLM execution (Gemma 4 / LiteRT) with sandboxed operating system tools.
• Resilient SaaS Architecture: Multi-tenant session isolation, reverse-engineered real-time protocols (WebSockets), and debounced persistence layers.

[llm-ops multi-provider-fallback pwa zero-cost-infra]

A fully automated hyper-local news platform that scrapes, classifies, translates via AI, generates graphics, and publishes without any human intervention.

• Orchestration Engine: Implements a cascading model fallback chain (Gemini primary, Groq/Kimi backup) with automatic credential rotation.
• GIST Diversity Algorithm: Leverages embedding-based similarity checks (arXiv:2405.18754) to detect and filter out duplicate news stories from overlapping feeds.
• Media Compositing: Multi-stage image generation pipeline featuring AI imagery, curated stock libraries, and programmatic SVG-canvas composite card rendering with progressive quality degradation paths.
• Live Demo

[chrome-extension manifest-v3 context-detection hybrid-decision-engine]

A high-performance Chrome Extension designed to automate document processing and resize images dynamically based on target form requirements.

• Sherlock Scanner: Deep context detector that parses table layouts, row hierarchies, and input field rules dynamically from the webpage.
• Polymorphic Resizer: A local client-side image processing engine that enforces strict aspect ratios (350x450 for Passport photos, 300x100 for signatures).
• Hybrid Decision Engine: Blends explicit page-level text dimensions with pre-defined profiles using a Safe Upscaling Guard to prevent image blur.
• Safety Buffer Logic: Computes target file sizes with precise buffer margins (Limit - 10KB if >50KB, Limit - 3KB if <=50KB) to ensure compatibility.
• Love Mode (Non-Blocking Concurrency): Decoupled UI and processing state machines, allowing concurrent multi-file compression without blocking.

[flutter dart gemma-4 on-device-ai privacy-first]

A mobile/desktop autonomous agent executing entirely on-device using Google's Gemma 4 (LiteRT) to ensure total data privacy.

• Sandboxed OS Tools: Equip the agent with 7 tools (Bash execution, file operations, local notifications, audio playback).
• Omega Fortress Security: Features custom protection modules including SentryPurity (command injection guard) and PathJailer (symlink traversal blocker).
• Context Compressors: 7-layer context manager that uses local summarization to run long conversations within tight hardware context limits.

[nodejs baileys websockets biological-state-machine]

A high-throughput Node.js microservice managing active real-time messaging sessions with contextual AI.

• Biological State Machine (BioClock): Introduces a per-session energy budget (0-100) that decays with activity and recharges over time. Features an "adrenaline override" when sales intent is detected.
• Adversarial Anti-Loop Guard: Appends invisible Unicode markers to outbound messages to prevent infinite bot-to-bot chat loops on shared channels.
• Resilient Infrastructure: Reverse-engineered WebSocket reconnects with exponential backoff and jitter, staggered session startups, and debounced database persistence.

[python gemini-1.5 cognitive-core system-automation]

A localized digital assistant using Gemini 1.5 Flash to act as a system-level agent.

• Cognitive Memory Core: Long-term JSON-based state storage allowing the assistant to remember user preferences across sessions.
• System Actions: Controls OS functions, application launching, and browser navigation via a conversation interface.

• LinkedIn: Abhishek Khuriwal
• Portfolio: abhikhrwl.github.io/my_portfolio
• Email: dev.abhishek.ai@gmail.com
• YouTube: @Aurelio_Tech-o4u

"Elite systems don't require enterprise budgets — they require better architectural decisions."