IIT Roorkee · Production & Industrial Engineering
I build agentic AI systems and data products, and measure whether they actually work.
lokesh = AIEngineer(
name = "Lokesh Kumar",
building = ["Travel-OS — AI operating system for travel agencies",
"Khata-AI — WhatsApp payments to a Google Sheets ledger (LangGraph + MCP)",
"A human-approved customer-support agent with its own eval platform"],
superpower= "turning LLM agents + data into working, measured products",
stack = ["Python", "LangGraph", "MCP", "RAG", "FastAPI", "PyTorch"],
)| Project | What it does |
|---|---|
| Khata-AI | LangGraph + MCP pipeline that turns WhatsApp payment messages into a Google Sheets ledger. Eval-driven: 94.8% → 100% held-out field accuracy |
| Travel-OS | AI operating system for travel agencies: multi-tenant CRM, itinerary & creative agents, human-in-the-loop approvals |
| Agentic-AI-Customer-Support-System | Support agent where the model proposes and code decides: no refund without human approval, plus an evaluation platform |
| delivery-network-intelligence | Graph-ML ETA model on 144,867 Delhivery shipments: MAE 107 → 32 min, ranked bottleneck hubs. Live demo |
| EV-dynamic-tariff-optimization | Agentic dynamic pricing for EV charging: simulated +8.2% revenue and −22% peak wait vs flat Rs15/kWh |
| decoding-customer-value | SQL + Power BI segmentation of 3,900 D2C customers; 43% of revenue is promo-reliant, with a promo-sunset plan |
| rlhf | SFT → reward model → PPO on Qwen2.5-0.5B with TRL; PPO beats SFT in 57.7% of reward-model comparisons, with an audit of its failure modes |
| RAG_Langchain | RAG over PDFs with LangChain, evaluated with Giskard |
Languages
ML / Deep Learning
LLMs / Agents
Backend / MLOps / Cloud
Data Stores
- 🏔️ Deputy Secretary · Himalayan Explorers' Club — led a 13-member expedition to 5,500m with zero incidents
- 🏐 Secretary & Captain · Volleyball — IIT Roorkee, 3 Inter-IIT medals
- 🎭 Founding Member · Rajasthan Cultural Association — events for 400+ people
