Manish S.

If your AI feature worked great in the demo and got quietly ignored two weeks after launch, I can probably tell you why before you explain what it does. It's rarely the model. After enough builds — and enough rescues — the pattern repeats: Retrieval was an afterthought. Nobody designed how the AI finds truth in your data, so it guesses. Confidently. Users notice within a week and quietly stop asking it anything real. Permissions never got designed. The agent has access to everything, so the first security review kills it. This one's avoidable, and almost nobody avoids it. Nothing gets measured. No evals, no observability. When it breaks, nobody can say what changed. When it improves, nobody can prove it. I build that layer — retrieval, permissions, evaluation — and the product around it. Get it right the first time, or rebuild it right the second. Most weeks, it's both. I'm Manish, founder of MDev Solutions. Here's what "getting permissions right" actually looks like in practice: a client wanted an agent with full write access to their CRM. We scoped it to three specific actions instead — narrower access, same outcome. Fourteen tools, role-based, every call logged. Zero data leakage, six months in production. That one decision is usually the difference between a security review that takes a day and one that kills the project. Two more, for context: → Knowledge assistant over 50,000+ documents: 92% accuracy, citation-backed answers, 68% of Tier-1 support tickets deflected. → AI-powered B2B platform scaled to 1M+ users. Voice agents on live inbound calls. Integrations writing to Xero, Salesforce, HubSpot — access controls security teams approve instead of veto. Stack: Claude & OpenAI APIs · MCP · RAG (hybrid search, pgvector, reranking) · LangGraph, CrewAI · Voice AI (VAPI, Twilio, Deepgram, ElevenLabs) · Python, Node, Next.js · PostgreSQL · AWS, GCP · n8n If you've got an existing codebase — even a messy one — that's fine. I inherit other people's code regularly and tell you honestly what's salvageable. Usually more than you fear. What I don't do: GPT wrappers, demo-ware for pitch decks, features bolted on because the board asked. I'll say so in the first message. Sometimes the right answer is "don't build this." You'll hear that early, not after three invoices. Working together: • Partner-CTO Retainer — weekly cadence, roadmap ownership • Fixed-Scope Build — clear SOW, milestone-based • Architecture & Advisory — audits, decision memos Several clients who started with a single fixed build are now on retainer. That's the actual measure of whether this works. Top Rated Plus · 100% Job Success · 4.9 on Clutch · 25+ clients · 10+ countries Response within 0–4 hours. Comfortable across US, EU, and ANZ time zones. Helping businesses become AI-first — an operating state, not a slogan. Send me what you're building, or what's broken. First call is a 30-minute architecture conversation. You'll leave with a scope, or a shorter plan than you came with.