Harshavardhan R
Focused builder,leveling up daily,turning goals into result.
New to Contra
Harshavardhan is ready for their next project!
Supermarket Automation System — End-to-End Retail AI Pipeline
Built a full automation system for a supermarket client handling 30,000+ SKUs — demand forecasting, WhatsApp ordering, and ERP integration, all wired together without manual intervention.
What I built:
WhatsApp Business API integration for automated customer ordering
Gofrugal ERP sync for real-time inventory and order management
HDBSCAN clustering for SKU segmentation and product grouping
Prophet time-series model for demand forecasting across 30k+ SKUs
n8n workflow automation connecting all systems end-to-end
The hard part:
30,000+ SKUs is not a toy dataset. Getting Prophet to forecast reliably across that many products, with seasonality and irregular demand patterns, while keeping the pipeline fast enough to be operationally useful — that was the real engineering challenge.
Impact:
Reduced manual ordering effort, improved stock prediction accuracy, and gave the client a WhatsApp-native interface their team could actually use without technical training.
0
3
SalesAutoPilot — AI Multi-Agent Sales Automation (Top 23, Supervity Hackathon)
Built a 10-agent AI system that fully automates the sales workflow — from morning briefings to inbound leads to client calls — placed top 23 nationally in India.
What I built:
10 specialized agents + 1 orchestrator managing the full sales cycle
4 automated flows: Morning Brief, Inbound Lead, Existing Client, STT Call Upload
Built entirely on Supervity's agentic platform
Designed, built, and deployed within hackathon timeframe
The hard part:
Most hackathon teams build one agent that does everything. The real challenge was decomposing the sales process into distinct responsibilities — each agent owning one job and doing it well — then orchestrating them without conflicts or redundancy.
Result:
Top 23 out of national pool of participants. Judges gave positive feedback with zero negative criticism on the system design.
0
9
Orchestara — Autonomous Accounts Payable Orchestration Engine
Built a production-grade multi-agent AI system that fully automates the Accounts Payable workflow — from invoice intake to payment execution — with zero manual intervention.
What I built:
14-stage deterministic state machine covering the full AP lifecycle
LangGraph Supervisor-Worker multi-agent architecture
Redis Streams for real-time event processing
3-layer LLM validation to ensure decision accuracy
6-signal fraud detection engine
Immutable audit trails for compliance
Published research at IJSART
The hard part:
Anyone can chain a few LLM calls together. The real challenge was making the system reliable — building validation layers that catch LLM hallucinations before they affect financial decisions, and designing a state machine that handles edge cases without human intervention.
Core philosophy:
LLMs are workers, not decision makers. Every critical decision in Orchestara goes through deterministic logic first — the LLM handles understanding, the system handles judgment.
Impact:
Reduces AP processing time and human error in enterprise finance workflows. Designed to handle real invoice volumes at scale.
1
14
O2C Graph Intelligence — Natural Language SAP Query System
Built a natural language interface over SAP's Order-to-Cash process using an 836-node knowledge graph. Users ask business questions in plain English and get instant answers — no SQL knowledge needed.
What I built:
836-node SAP O2C knowledge graph mapping the full order-to-cash domain
NL→SQL pipeline using Groq + LLaMA 3.3-70B
Live deployed web interface for real-time querying
The hard part:
SAP O2C is a dense enterprise domain — invoicing, collections, order fulfillment, payments. The challenge wasn't just building the LLM pipeline, it was modeling the domain correctly so the queries actually make business sense.
Impact:
Makes complex SAP data accessible to non-technical business users without any SQL or SAP expertise.
1
18