Venkata Manideep Patibandla
AI Engineer · Agent Systems · ML Engineering · Production AI
New Haven, CT
I’m an AI engineer. In plain terms, I build AI systems that don’t just look good in a demo — they hold up when real users and real money are on the line. Most of my work comes down to one idea: making AI reliable enough to actually trust in production.
A few things I’ve built:
RelayOps — a customer-support agent for a telecom company that can actually do things (look up an account, check a plan) instead of just chatting. The hard part isn’t the conversation — it’s the guardrails: I built it so it can’t invent fake offers, leak someone’s personal information, or take an unsafe action, then stress-tested it with tricky, adversarial cases to prove it stays in line. Try the live demo.
CostGuard — think of it as a smart switchboard that sits between an app and the AI models it relies on. It quietly checks each AI response for quality, automatically retries or switches to a different model when one falls short, and tracks exactly what every call costs. In testing it cut AI spending by 10–20× versus the obvious default — same quality, a fraction of the bill.
RealDataAgentBench — to make CostGuard’s model-switching smart, I first had to answer a simple but expensive question: which AI model should you actually use, and what will it cost? So I ran 1,412+ head-to-head tests across 12 of the leading AI models on real data tasks. The headline finding: a small, cheap model held its own against the top-ranked one while costing 65× less per task — exactly the kind of result that saves a company real money.
Tether (in development) — a safety net for AI agents that run long tasks, so that if one crashes or a provider goes down halfway through, it can pick up where it left off instead of starting over.
These days I’m an AI/ML Engineer (Contract) at a stealth AI startup, leading a model that estimates what a home is worth from just its address — accurately enough (within about 5%) that lenders can skip a manual appraisal on standard properties, and roughly half as far off as the data provider’s own tool. Alongside that, as a Forward Deployed AI Engineer at SBL, I work hands-on with companies (Emoha, Topmate, Clinik) to automate their sales and operations workflows from start to finish. Earlier, at AnternData Solutions, I delivered AI projects for global tech companies including TigerData (formerly Timescale). I’m also IBM Certified in Agentic AI.
news
| Jun 7, 2026 | Released RelayOps v1 — a production-shaped telecom support agent with scoped tools, guardrails, RAG citations, adversarial evals, a Qwen LoRA intent classifier, and a live Railway demo. |
|---|---|
| Apr 1, 2026 | Started as an AI/ML Engineer (Contract) at a stealth AI startup, leading an automated valuation model (AVM) from inception to production — 5.1% MAPE, valuing homes within ±5% of appraised value on live API data. |
| May 15, 2025 | Graduated with an M.S. in Computer Science from Sacred Heart University (GPA: 3.8/4.0). Inducted into the Upsilon Pi Epsilon (UPE) Honor Society. |
| Apr 11, 2025 | Released RealDataAgentBench — an open-source benchmark across 12 frontier LLMs and 1,412+ runs surfacing the correctness vs. statistical-validity gap in frontier models. |
| Apr 1, 2024 | Joined SBL as a Forward Deployed AI Engineer, deploying RevOps and workflow-automation solutions on SBL’s Konduit platform for enterprise and growth-stage clients (Emoha, Topmate, Clinik). |
latest posts
| May 19, 2025 | How to Actually Use Claude: 18 Steps That Unlock 100% of Its Potential |
|---|---|
| May 14, 2025 | The Most Expensive Mistake in LLM Engineering (And How to Fix It With Data) |
| May 10, 2025 | KV Caching in LLMs |