Product-minded AI systems builder focused on production RAG, agentic workflows, LLMOps evaluation, and data products that move retention, cost, latency, and risk metrics.
Risk Impact
$300M+ surfaced
Automation
500+ tickets / week
Growth
6x engagement lift
Hybrid RAG + relational search over 1,000+ supplier contracts, designed with query rewriting, source routing, sub-query decomposition, and re-ranking to surface liability without hallucinating in high-stakes decisions.
Multi-agent content workflow for GTM teams: scraper, grader, RAG validator, LLM generator, and QA loop with stateful handoffs to raise content pass rate toward 85% while controlling hallucination risk.
CI-style evaluation layer for RAG and agent systems: malicious-prompt stress tests, faithfulness scoring, answer relevance, context precision, and dashboard exports so model behavior is tested like production software.
Build track for privacy-sensitive domain models: fine-tune Llama or Mistral on niche documentation with QLoRA, track loss and GPU memory, then quantize to 4-bit for low-latency local inference.
Streamlit investment intelligence dashboard that turns historical VC data into thesis-ready insights—exit efficiency, geographic capital density, sector deep dives, investor matchmaking, and a Random Forest startup success predictor.
End-to-end production ML system that predicts EV charging station surge risk: reproducible data prep, shared feature engineering, validation gates, drift monitoring, FastAPI inference with Prometheus metrics, Docker, and GitHub Actions CI/CD—runnable with no external datasets or cloud accounts.
Streamlit app for large-scale ChatGPT app review analysis: upload a CSV, run LDA and NMF topic models, view aggregated theme distributions, and export predictions—with models trained in Colab and tracked in Databricks MLflow.
Hybrid retrieval with sub-query decomposition, BM25+vector fusion, and BGE re-ranking. Built to surface facts without hallucinating on high-stakes decisions.
Stateful multi-agent pipelines with typed handoffs, human-in-the-loop checkpoints, and graceful fallbacks.
CI-style eval harness: faithfulness, context precision, answer relevance, malicious-prompt stress testing, and drift monitoring.
Executive decision dashboards, risk-scoring pipelines, and dbt data models that power $300M+ in liability visibility.
Fine-tune Llama/Mistral on proprietary docs with QLoRA, then quantize to 4-bit for sub-100ms local inference.
Behavioral cohort analysis, funnel modeling, and LLM-assisted VoC analysis over 1M+ app reviews.
"I build AI products where model behavior is measured against business outcomes: liability exposure, support latency, engagement lift, and hallucination risk."
AI that can't be measured doesn't survive production — evaluation and business metrics come first, benchmarks second.
— Product + AI Systems Operating Principle
Mokshith Kumar builds at the intersection of product strategy, data systems, and applied AI. His work spans enterprise contract intelligence, supplier automation, AI-driven lifecycle messaging, hybrid RAG support, agentic RevOps, and large-scale voice-of-customer analysis over roughly one million app reviews.
Arizona State University · W. P. Carey
Blip · AI notifications, RAG support, retention analytics
Cisco · contract intelligence, supplier workflows, $300M+ liability visibility
Cisco · Snowflake dashboards, risk scoring, executive decision systems
Open To Competitive AI Roles
Targeting AI Product Manager, AI Application Developer, and Full-Stack AI Engineer roles where product sense and production AI depth both matter.
Contact Mokshith →