Staff AI/ML Engineer · 8+ yrs · online

I build the system layer that makes AI products actually work.

I'm Kaushal. I design and ship the unglamorous layer most AI demos skip —agent runtimes, harnesses, evals, and developer platforms — the infrastructure that makes AI reliable, safe, and easy to build on at scale.

Read the writing →About me
FIG.01 — UNIFIED AI RUNTIMEwhat I build
15M+conv / quarterConversations & 800B+ tokens per quarter on the AI runtime I architected at HighLevel
600+operationsExposed over MCP in the headless developer platform I built — mcp · sdk · cli · plugins
4.7M+voice callsReal-time voice conversations served at 99.9%+ success
3publicationsPeer-reviewed: EMNLP 2024 · Elsevier ASOC ×2
SEC::CONTEXT

What I work on

Currently at HighLevel, where I architected the unified AI runtime and agent harness behind 6+ product surfaces (Voice AI, Conversation AI, Agent Studio, Ask AI, Ads Copilot, Email Agent). Previously: compliance agents at Zania (evidence retrieval F1@10 86% → 97%) and a unified Document AI stack at Intellect (replaced 23+ models, cut OCR spend 70%). I care about the boundary where ML research meets production systems — model routing, tool calling, long-context state, memory, eval gates, and the reliability engineering that keeps it all up.

SEC::MODULESall writing →
MOD::AGENTSAgentic SystemsAgent runtimes, harnesses, tool calling, memory, and orchestration.MOD::EVALSEvals & ReliabilityRelease gates, faithfulness, and the reliability engineering that keeps AI up.MOD::RAGRetrieval & RAGHybrid retrieval, reranking, and citation-grounded answers.MOD::SYSTEMSML SystemsModel routing, serving, cost engineering, and Document AI at scale.MOD::PAPERSPapers & ResearchNotes on deep-learning research and my own peer-reviewed work.
SEC::WRITINGall posts →
01The system layer behind AI products: what an agent runtime actually does
· 4m
02From 86% to 97%: engineering evidence retrieval for compliance agents
· 3m
03One model instead of 23: consolidating a Document AI stack
· 3m
SEC::RESEARCHall papers →
3 PEER-REVIEWED PUBLICATIONS

EMNLP 2024 · Elsevier Applied Soft Computing ×2 — on Document AI, de-identification, and parameter-efficient fine-tuning.

Read the papers →
SEC::CONTACT

Hiring for Staff / Principal AI platform work?

I write here about building production AI systems. If that's your world too — or you're building an AI platform team — let's talk.

LinkedIn →EmailGitHub