Yadnyesh Mulay: I build useful things
and ship them with ai.
Full-stack product builds, AI integration, business/AI transformation consulting, and hands-on teaching for students and professionals.
Your idea. your users.
Your product.
I take ideas from a first sketch to a working product, building the full stack, weaving in AI where it earns its place, and helping people grow along the way.
Build
Full-stack apps, MVPs & prototypes
Enhance
LLM integration, RAG & automation
Consult
AI transformation roadmaps & process optimization
Teach
Mentoring students & upskilling professionals
Where I've worked.
MahaWiki: Tech Community
Running free, industry-oriented training, webinars, and workshops to make technical education accessible.
Cloud Community Group
Organising cloud events, bootcamps, and mentorship for developers and students.
SkillShip Foundation
Led expert sessions and skill-development workshops across colleges and companies.
The real index
Every entry names its own role and tech honestly, no invented metrics, no borrowed logos. Repositories verified under github.com/MYadnyesh.
- 04Twitter Sentiment Detector (Capstone)Source
T-Shaped Engineering
Vertical depth in AI/LLMs. Horizontal breadth across Cloud, Full Stack, Product/UX, and Business Consulting, resting on software engineering fundamentals.
AI / LLMs
Deep specialization in LLM application development: prompt engineering, RAG systems, agentic workflows, multi-modal AI, and AI-native product architecture.
The full toolkit
A wall of what I build with, grouped by where it sits in the stack. Hover any row to slow it down.
AI Systems Engineering
AI isn't a bolt-on feature, it's an integrated engineering layer, from API call to production workflow.
LLM APIs
Integration & orchestration
- LLM APIs (OpenAI, Gemini, Claude)
- Prompt Engineering
Retrieval
Grounding & context
- RAG & Embeddings
- AI Automation
Agents & Workflows
Autonomous reasoning & orchestration
- AI Agents
- Multimodal AI
- Voice AI
- Agentic Workflows
Infrastructure
Local inference & serving
- Ollama / Local LLMs
- MCP (Model Context Protocol)
- OpenCode / Codex / OpenClaw
- LangGraph / LangChain
- CrewAI / Multi-Agent
- Coding Agents
- Model Routing
- Inference / Quantization
- Fine-tuning / LoRA
- AI Infrastructure
- Voice Agents
- n8n / Workflow Automation
AI is a capability, not the product
LLMs enable better products, they don't replace product thinking. The value is in what the system enables users to do.
Evals before vibes
Systematic evaluation beats prompt tweaking. Build eval infrastructure first.
Ground in reality
RAG, citations, and verification aren't optional. Hallucination is a feature of the architecture, not a bug to prompt away.
Design for the failure modes
Latency, cost, rate limits, degradation. Graceful degradation beats a perfect happy path.
Human + AI beats AI alone
Best systems keep humans in the loop for judgment, creativity, and accountability.
Start simple, scale deliberately
Single LLM call → RAG → Agent → Multi-agent. Add complexity only when the simpler layer is proven insufficient.
Inpublic
The visible layer: shipped projects, open-source repositories, community work, and teaching. Everything you can check before we ever talk.
In thelab
The experimental layer: local LLMs, agentic workflows, MCP, multi-agent systems, and inference. Exploration that feeds the next thing I build.
In the lab
Technical directions I'm investing time to understand deeply, ahead of production use. Deliberate exploration, not unfinished work.
Local LLMs & Inference
Running models locally with Ollama, experimenting with quantization (GGUF), model routing, and offline-first AI architectures.
Agentic Coding Workflows
Evaluating OpenCode, Codex, OpenClaw, and MCP for autonomous code generation, refactoring, and multi-file edits.
Multi-Agent Systems
Building collaborative agent networks with LangGraph and CrewAI, planning, delegation, and verification loops.
AI Infrastructure & Serving
Model serving (vLLM, TGI), batch inference optimization, KV caching, and cost-effective deployment patterns.
Voice AI & Real-time Agents
Speech-to-text, LLM reasoning, text-to-speech pipelines with latency optimization for conversational agents.
Workflow Automation (n8n)
Visual workflow builder for AI-augmented business processes, connecting APIs, databases, and LLMs without code.
Exploration is not distraction. These areas are chosen deliberately, based on where AI-native software is heading.
Latest writing
What I have been writing and posting lately.
The model is rented. The harness is what you build.
Building reliable AI agents takes more than a good model. It takes a system that controls what the model can access, what it can do, and how its output gets verified. That system is the harness: the tools, context, memory, permissions, and controls around the model. The model is rented. The harness is what you build. This changes how I approach debugging. Instead of asking, “Why did the model fail?”, I’d ask: Which layer failed? Was the instruction unclear? Was context missing? Did a tool fail? Were permissions too broad? Did nobody verify the result? Each failure points to a different fix. Anthropic's work on long-running agents shows what this looks like in practice: track hundreds of features, build them incrementally, and verify each one end to end before calling it done. Replit’s July 2025 database incident shows the other side. The agent deleted a production database despite an explicit code freeze. The rule existed. The enforcement didn’t. That’s why my baseline for agentic systems is simple: ->A written spec ->Scoped permissions and a sandbox ->Verification before completion ->Human approval for destructive actions ->A rollback path if things go wrong The spec defines the contract. The harness makes it enforceable. What’s the first layer you investigate when an agent misbehaves? Checkout at www.yadnyesh.dev #AIAgents #AIEngineering #SoftwareEngineering #ArtificialIntelligence #LLM #GenerativeAI #AgenticAI #LLMOps #AIInfrastructure #DevTools #SoftwareArchitecture
Evals Before Vibes
OpenAI's Dots demos stalled on stage at DevDay, in front of 2,500 people. For anyone building software, that's a familiar problem: getting something to work in a demo is one thing. Making it reliable in production is another. That's where evals come in. Define what "working" means. Test it. Run those tests whenever you change a prompt, model, or workflow. Anthropic shared an example where a prompt change pushed success rates from 30% to 80%. They caught the improvement because they already had evals in place. Without that feedback loop, you're shipping changes and hoping for the best. We've seen what happens when AI gets things wrong. Air Canada's chatbot invented a bereavement policy, and the airline was held responsible. Klarna also had to walk back AI-generated outputs after accuracy issues surfaced publicly. Full breakdown in the carousel. Checkout www.yadnyesh.dev #AIEngineering #GenAI #ArtificialIntelligence #MachineLearning #SoftwareEngineering #EVALS
The spec is the product.
Everyone’s arguing about which AI agent writes the best code. Wrong debate. As builders, the harder problem isn’t getting an agent to write code. It’s giving it a problem that’s clear enough to solve. Instead of: “Build a login page.” I’d specify: “Build email/password login with validation, loading and error states. Disable submit until valid. Redirect successful logins to /dashboard. Never log passwords.” Now the agent has constraints to work with and I have something concrete to verify. That’s the shift: You define the problem. The agent explores the solution. The better the spec, the fewer cycles you spend fixing misunderstandings. Prompts are temporary. Specs is the product. Full breakdown in the carousel. Checkout www.yadnyesh.dev #AI #SoftwareEngineering #Leadership #Innovation #FutureOfWork
Yadnyesh Mulay, AI-first full-stack developer and business/AI consultant. I build useful software and advise on AI-driven business transformation. I turn ideas into working technical prototypes, and turn processes into AI & agentic AI roadmaps.
AI-first T-shaped Full Stack Developer & Business/AI Consultant, comfortable moving between product decisions, system architecture, the model layer itself, and the business process it all needs to serve. This is the record of how that came together.
Full-stack foundations
Shipped web applications end-to-end, React, Node.js, databases, deployment.
MSc Computer Science, University of Greenwich
Distinction. Capstone: Track-Master, a MERN train-ticket booking system with TfL API integration, PDF ticket generation, and payment processing.
Cloud & platform certification
AWS Solutions Architect – Associate, Azure Fundamentals (AZ-900), Azure Developer Associate (AZ-204), Google AI Essentials.
AI-native software
OutSkill AI Generalist Accelerator. Built Prism, the Microsoft 365 Sentiment Analyzer, and the Gmail/Chat Sentiment Analyzer, putting LLM APIs, RAG, and hybrid NLP into working products.
AI & business consulting
Advising on AI & agentic AI transformation roadmaps, technical product and program delivery, business process optimization, and no-code/low-code workflow automation (vibe coding).
AWS Solutions Architect – Associate
Amazon Web Services · 2024
Microsoft Azure Fundamentals (AZ-900)
Microsoft · 2023
Microsoft Azure Developer Associate (AZ-204)
Microsoft · 2024
Google AI Essentials
Google · 2024
OutSkill AI Generalist Accelerator
OutSkill · 2024
Skillship Foundation Aatmanirbhar Program, Best Performer, Python
Skillship Foundation · 2022
Community Manager
MahaWiki · 2022–Present
Open knowledge contribution and community coordination.
Community Manager
Cloud Community Group · May 2021 – Present
Cloud-native technologies, Kubernetes, serverless patterns.
Campus Club Lead
GESCOE MOZCLUB / Mozilla Campus Club · 2019–2021
Privacy/security advocacy, Firefox Add-ons development, HTML/WordPress workshops.
Firefox Beta Tester
Mozilla India, Quality Assurance · 2017
Volunteer QA contributor for Firefox Developer Edition test days (Preferences Search, CSS Grid Inspector, Form Autofill) and the Firefox Quantum website-compatibility testing push.
Campus Ambassador & Best Performer (Python)
SkillShip Foundation · 2020–2022
Ran student webinars and skill-building programs; recognised as Top Active Member at HackOnfest 2020 and Best Performer in the Aatmanirbhar Python track.
Next step
Start with the real idea.
Tell me what you're building or trying to learn. I'll tell you if I can help.
Let's build something useful.
I work with founders, engineering teams, and organizations building AI-native products and full-stack applications, and with business leaders shaping AI & agentic AI transformation roadmaps, whether that's a technical prototype, an AI integration, a process optimization engagement, or a full MVP.
Build
MVP Development · Full-Stack Applications · Web Applications · Technical Prototypes
Enhance Intelligence
LLM Integration · AI Automation · Agentic Workflows · Retrieval Systems (RAG)
Architect
System Design · Technology Selection · Scalability Planning · API Strategy
Prototype
Idea → Working Prototype · Feasibility Validation · Technical Spike · Proof of Concept
Product
UX + AI-Native Strategy · Product Discovery · Feature Prioritization · Metrics Design
Explore
Emerging Tech Feasibility · Local LLM Deployment · Agent Architecture · AI Infrastructure
Consult
AI & Agentic AI Transformation Roadmaps · Technical Product & Program Delivery · Business Process Optimization · No-Code / Low-Code Workflow Automation (Vibe Coding)
Form submissions are used only to schedule a conversation, not stored in databases, not shared, not used for marketing. No tracking pixels.
Rather just talk?
Grab a 30-minute slot on my calendar.
Rather just pick a time? Book a 30-minute call directly below, no message required.
Pick a time
Opens my live calendar with available 30-minute slots.