Ask Kiran
AI-powered assistant
Try asking:
Favourite Repos
Loading starred repositories…
Agentic Stack
My AI Stack
Hover any card to see details
User Interfaces
Where I interact with my agent stack
Primary mobile interface
Chat-based interaction with PA agent. Voice messages, inline commands, media delivery. Always-on connection via Telegram API.
Hermes CLI interface
Direct shell access. File ops, git, deployment, agent spawning. Full power-user control surface via macOS terminal.
Hermes Desktop app
GUI companion for Hermes Agent. Visual agent logs, file browsing, session management, config editing.
Agent Orchestration
The brains of the operation
Core AI agent framework
Open-source agent by Nous Research. Multi-profile, tool-calling, skill library, cron scheduling, cross-profile memory, and subagent delegation.
Personal Assistant
Life management agent: tasks, reminders, calendar, projects, wiki. Proactive suggestions, daily briefs, evening recaps, Obsidian vault master.
Dev · Creative · Content
Separate Hermes profiles for code, design, and content generation. Each with tailored skills, tool sets, and model preferences.
Skills & Tools
What the agents can do
30+ curated skills
Reusable workflows across 9 categories: media creation, software dev, research, social media, productivity, MLOps, design, and process.
Web · File · Terminal · API
Full tool belt: web search/extract, file read/write, shell execution, GitHub, browser automation, cron scheduling, subagent delegation.
Multi-bot infrastructure
Delivery channels for cron output, media delivery, inline commands. Home channel for scheduled results and notifications.
Memory & Knowledge
Persistent state and structured knowledge
Cross-session memory
Persistent memory system storing preferences, project state, environment facts. Survives across sessions and different agent profiles.
PA database & wiki
Structured markdown vault: tasks, projects, areas, daily notes, calendar, inbox. Single source of truth for the PA agent.
Karpathy-style knowledge base
Interlinked markdown wiki with entities, concepts, comparisons, and raw source material. Agent-readable knowledge graph.
Large Language Models
The reasoning engine
Primary reasoning model
Fast, powerful general-purpose LLM. Default model for PA and Hermes agents. Accessed via DeepSeek API with low-latency inference.
Secondary / fallback model
Alternative model for specialized tasks. Different strength profile. Used as fallback when primary is unavailable or for cost-sensitive cron jobs.
Infrastructure
The hardware running it all
Always-on home server
Primary host for Hermes Agent. Runs 24/7, manages cron jobs, hosts terminal sessions, runs background agents. macOS 12.7.6.
Mesh VPN network
WireGuard-based mesh VPN connecting all devices securely. MacBook, notebooks, phones — all on one network, accessible from anywhere.
Unified LLM gateway
Single API for multiple models. Routes to DeepSeek, MiMo, GPT, Claude, or any supported provider. Free-tier fallback for cron jobs.
About me
I build AI systems. I own the product.
I started in operations, building automated workflows with Microsoft Power Platform — Power Automate, Power Apps, SharePoint. Over time, those workflows evolved: from rule-based automation to n8n pipelines, from chatbots to autonomous agents. Today, I architect and ship production AI systems using MCP servers, RAG pipelines, and agent frameworks.
I don't believe in AI demos that never ship. I build working systems — agents that run 24/7, pipelines that publish without intervention, tools that solve real problems. No buzzwords. No slide decks. Working software.
What makes me different: I can lead a product from strategy through deployment. I understand the business case, the user journey, and the technical architecture — because I've owned all three. That's the rare combination I bring to every initiative.
10+
years shipping products
Automation → AI
Power Platform → n8n → Agentic AI
Full-stack
Agentic AI architecture
Credentials
Expertise
What I Bring
Product & Automation
PRODUCT OWNERSHIP
End-to-end product strategy, backlog management, roadmap delivery, and stakeholder alignment
WORKFLOW AUTOMATION
Autonomous workflows, no-code/low-code platforms, and API-driven orchestration
DATA & ANALYTICS
Dashboards, SQL optimization, reporting frameworks, and data-driven decision making
DEVELOPMENT STACK
Languages, testing, infrastructure, and engineering practices I ship with
PLATFORM EXPERIENCE
Banking, fintech and identity platforms I have shipped on and integrated with
AI & Builder Stack
AGENT ARCHITECTURE
Harness design, tool-use loops, reflection cycles, multi-agent orchestration and swarm topologies
RAG & RETRIEVAL SYSTEMS
End-to-end pipeline design — ingestion, chunking, embedding, hybrid search, reranking, generation
LOCAL LLM DEPLOYMENT
Model quantization, inference optimization, hardware-aware serving and benchmarking
CONTEXT & MEMORY
Context window management, token budget design, agent memory architectures (episodic, semantic, external)
MCP ARCHITECTURE
Custom MCP server design and production deployment, multi-server orchestration, tool/resource exposure
VOICE & MULTIMODAL PIPELINES
STT→LLM→TTS stack design, real-time streaming, latency optimization for voice AI systems
Projects
Things I've Built
Instant Newsletter Generator
A LangGraph-powered pipeline that researches trending topics, deduplicates, curates, formats, and styles a complete HTML newsletter — all in under a minute. SSE streaming shows every step live.
Five agent nodes chain together in real time: research → dedup → curate → format → style. Each node streams its progress via Server-Sent Events so you watch the newsletter assemble itself.
Toronto Knows LLM
A hyper-local fine-tuned language model that knows Toronto inside out — neighborhoods, slang, food, transit, culture, and community voices. Trained on Qwen3.5-4B using MLX QLoRA.
Every LLM gives generic 'visit the CN Tower' advice. Toronto Knows speaks with the actual voice of the city — real slang, hidden gems, multicultural perspective. No other LLM is this specifically trained on one city.
Agent Builders Pipeline
A daily newsletter on agentic AI — automated end-to-end. Research cron → format cron → Substack draft, all running on Hermes agents.
Two cron jobs chain together every morning: one researches trending stories, the other formats and publishes. Kiran reviews each draft in 30 seconds before it goes live.
RAG-Based AI Voice Agent
A voice-enabled AI assistant built on Retrieval-Augmented Generation — demonstrating end-to-end AI systems thinking, not just prompting.
This wasn't about writing better prompts. It was about architecting a full pipeline: ingestion, embedding, retrieval, generation, and voice output — all working together.
© 2026 Kiran Khutal. Built with Next.js & Tailwind CSS.