Problem: fixed-interval engine maintenance wastes healthy life and still misses early failures. Built: end-to-end pipeline on NASA C-MAPSS (FD001–FD004) condition-aware normalisation, causal rolling features, six model families under grouped 5-fold CV; alert thresholds tuned on out-of-fold predictions and frozen before test; Mondrian conformal intervals; Streamlit demo. Result: capped test RMSE 12.3–13.2 cycles on all four subsets; 100% of fails-within-30 engines detected with ≈50 cycles median warning lead (cost-aware policy). The project caught and quantified its own preprocessing bug — the fix moved FD004 from 23.9 → 13.2 RMSE. Stack: Python · scikit-learn · PyTorch · Streamlit

Problem: manual review of pipeline corrosion inspections misses critical defects and burns engineer hours. Built: LangGraph multi-agent triage (planner → research → executor → critic with loop-back), human-approval gate, work-order generation — fully offline with a deterministic mock LLM; 23-test suite. Result: missed-critical recall 0.85 → 0.98 on the synthetic validation benchmark. Stack: Python · LangGraph · Streamlit · Plotly

Problem: ordering across delivery platforms is manual and UI-bound. Built: chat-first ordering API + React UI + Playwright agent that drives Swiggy end-to-end — dish matching, cart via the app's own Redux handlers, address → payments page; PlatformAdapter pattern; never charges, hands over for manual pay. Result: proven on a live order through to the payments page (₹264), with honest fallbacks when a variant or payment method doesn't exist. Stack: TypeScript · React · Express · Playwright · Python

• Undergraduate research intern School of AI & Data Science, IIT Jodhpur: agentic AI & multi-agent systems.

• GitHub — @DeltaN04