i build AI systems that actually do something useful - not chatbots, not wrappers, not "chat with your PDF" demos.
when i'm not obsessing over retrieval pipelines and document intelligence systems, i'm probably:
- reverse-engineering why some RAG system hallucinated with full confidence
- doing outreach to companies that don't know they have an AI problem yet
- or building something at midnight that'll either be embarrassing or impressive by morning
🏢 Internal Policy Intelligence System
Employees ask HR the same questions every day. HR gets annoyed. The answer is always buried in some doc nobody reads. I built a RAG system that answers policy questions in seconds, with source citations and access controls. Governance-controlled, version-aware, zero hallucination tolerance.
Live demo available · Targeting regulated Indian companies · try it live
📄 Document Comparison & Compliance Engine (in progress)
Compliance teams manually read two versions of a 40-page contract to find what changed and whether it's a violation. That's genuinely someone's job. I'm building the system that does it in under a minute.
Architecture complete · Active build
before I got obsessed with applied systems, I did proper research:
-
HYDRAN — brain tumor detection model that outperformed all benchmark architectures. yes, it's a cool acronym, no, I will not apologise for that
Deep Learning · Medical Imaging · Research-grade -
XGBoost Stock Trading Pipeline — compared ML predictions vs rule-based technical analysis on Indian equities. spoiler: both lose sometimes
Quantitative ML · Indian Markets · Shipped -
30-Day ML Challenge — 15+ projects in 30 days. regression to clustering to from-scratch implementations. built to stop being scared of the fundamentals
15+ Projects · From Scratch · Completed -
Ask Krishna — AI assistant grounded in the Bhagavad Gita. full-stack SaaS with auth, streaming AI, and a deliberately minimal aesthetic. built over a weekend
Full-stack SaaS · Streaming AI · Live -
AURA — voice-first assistive AI for elderly and visually impaired users. processes speech and visual input to help users understand their surroundings
Multimodal · Assistive Tech · Voice-first -
MedWaste Guardian — multi-modal AI system that classifies biomedical waste via image, voice, or text and guides correct disposal based on actual legal regulations
Multimodal · Compliance-grounded · Shipped -
AI Client Acquisition System — end-to-end sales pipeline. finds leads, writes personalised outreach, sends it, follows up, makes cold calls with a human-like voice, updates the CRM. fully automated
n8n · LLM Orchestration · End-to-end -
Siri + n8n Personal Assistant — turned Siri from a basic voice assistant into a domain-specific AI agent that handles emails, calendar, and web research through voice
Voice Agent · n8n · Personal Automation -
AI Cold Calling System — fully automated outbound calling workflow. AI places calls, delivers a human-like pitch, handles live responses, and updates CRM records in real time
Voice AI · CRM Integration · Fully Automated
explore more about me → Portfolio
- final year BTech, AI & ML
- open to remote AI engineering roles and freelance work involving document intelligence, RAG systems, or workflow automation
- i write about what i build → hashnode.com/@ayoozzzz
- i talk about it too → youtube.com/@devayushrout
i use Antigravity as my IDE because it talks back.
my rubber duck debugger has opinions now and honestly it's an improvement.