Pushpak

AI Agent Engineer and Data Scientist with 1 year of internship experience designing agentic AI workflows, prompt engineering multi-agent systems, and deploying end-to-end ML pipelines on AWS. Shipped a 5-agent Demand Planning Platform (Supervisor → Planner → Executor hierarchy) and production demand forecasting system achieving 26.1% WMAPE for a large-scale electronics manufacturer. Proficient in Agnetic AI Workflows - LLM orchestration (LangChain, LangGraph, MCP Servers, ReAct), AWS MLOps (SageMaker, Airflow, MLflow), and LLM fine-tuning (LoRA, Unsloth) — with a track record of delivering both rapid POCs and hardened production systems.

GenAI Intern — AI Engineer at Karini AI (2025-12 – Present)

• Designed the Bajaj Demand Planning Platform: a 5-agent hierarchical system (Demand Planning Supervisor → Execution Planner Agent → Analysis / Demand Sensing / SQL Agents) for NL business analytics on AWS Athena-backed trade and e-commerce data.
• Engineered structured JSON execution-plan prompts for the Planner Agent (target_agent, resolved_scope, steps, must_not bans) to prevent hallucination in downstream Haiku-class executor models — core prompt engineering work that enforced output contracts.
• Built the Demand Sensing Agent: executes 1–4 week short-term demand waterfall via a custom DS-Waterfall Lambda (DemandSensingAthena); integrated Visual Crossing weather API, Lookup DB, and Tavily MCP as tools; outputs waterfall-adjusted daily demand signals.
• Built SQL Agent for NL-to-Athena/Trino SQL translation across 3 complex intents (forecast_accuracy, period_compare, ds_vs_actual)
• Analysis Agent queries Athena via Lookup KB and returns WMAPE, Accuracy, and Bias by Business Unit in under 4 seconds.
• Built Demand Sensing Scheduler: daily webhook agent that auto-resolves date, fetches ASP filename via Lambda (getFilenameDemandSensing), triggers Airflow demand-sensing DAG with dynamic window parameter, and dispatches SES email notifications — enabling zero-touch daily runs.
• Bajaj Demand Forecasting: two-stage XGBoost (Classifier → Regressor) with 69 features across 8 layers (festival calendars, 5-tier ASP waterfall, lag/rolling stats, SKU lifecycle)
• Rolling Origin CV; conformal P10/P50/P90 bands; achieved 26.1% Trade WMAPE on 964K+ transactions — 22 pp below naïve baseline.
• Built ML Pipeline Control Agents (Chat, Analysis, Webhook) on Karini AI using ReAct and Tool Calling
• DynamoDB state tracking; async webhook pattern for SageMaker long-job completion; 94% intent accuracy across 8 conversation flows, 60 UAT test cases.
• Automated full MLOps lifecycle via SageMaker Pipelines (5-step training + 3-step inference), MLflow experiment tracking, conditional WMAPE quality gate → SageMaker Model Registry, and 9 Airflow EKS DAGs — cutting 4–6 hr manual cycles to under 40 min.
• Badger Load Optimization: Orchestrator + 4-Lambda pipeline (Data Prep → BFD POD Packer → Pass-2 Consolidation → LIFO Truck Assign → CSV Export with presigned S3 URLs) for a last-mile liquor client
• Orchestrator retries on failure — zero manual dispatch effort.
• SAP Invoice Query Agent: Karini AI + SAP HANA MCP (list_tables, get_schema, execute_query); auto-detects invoices overdue past Net 30, formats structured AR report, dispatches SES email alerts. PSEG Meter-to-Cash POC: zero-consumption exception-handling agent — system prompt, tool schemas, escalation logic.

Software Intern at CDK Global (2024-11 – 2025-06)

• Built CDK Service Book — real-time mobile app (React Native CLI, Redux Toolkit, SignalR,.NET Core Web API, EF Core) deployed to 16+ North American automotive dealerships for live service status tracking and technician workflow management.
• Applied Clean Architecture and Test-Driven Development (TDD) across.NET Core backend services; collaborated in a 5-member Agile team on enterprise automotive dealership platform features.

AI Intern at GenAI Kits (2024-06 – 2024-10)

• Developed LangChain + OpenAI chatbots and RAG pipelines (FAISS, Weaviate vector databases) for semantic search over domain-specific document corpora; integrated LLM-based agents into production FastAPI backends.
• Optimised prompt engineering, retrieval chunking, and embedding strategies to improve RAG answer relevance; worked in a 5-member team delivering end-to-end GenAI pipeline features.

B.Tech in Computer Science Engineering (AI & ML) – Vishwakarma Institute of Technology (VIT) (2022 – 2026)