Himanshu Laddhad

Building a single EDI-to-XML mapping by hand took three to five weeks. I architected a retrieval-grounded agent to collapse that, then trained a small local model to remove the per-call API dependency at production volume, and measured it honestly against the frontier-model pilot rather than claiming parity.

• Architected a RAG-grounded EDI-to-XML mapping agent routing messages through Groq to a GPT-4.1 tool-calling model, retrieving from a 400+ file XSLT corpus keyed by company and function, cutting mapping build time from 3–5 weeks of manual development to 3–4 hours.
• Trained a small language model (Qwen) with QLoRA covering transform generation, modification, repair and explanation, grounded by the same retrieval layer, reaching 73% first-pass Altova validity against the GPT-4.1 pilot's 78%, removing per-call API dependency across the full task set.
• Delivered the Streamlit evaluation harness fronting both pipelines, executing generated XSLT against user-supplied test cases in-session, handed off as the reference implementation for PartnerLinQ's engineering team to productionise.

The mandate was open-ended: modernise how the business used its operational data, with no predefined scope. I drew the boundary myself: my work began at the warehouse layer, consuming what upstream systems produced and holding them to a data contract rather than claiming ownership I could not enforce. The decision I would defend hardest is that failed rows are quarantined with a typed exception code, never dropped, because a missing record is not a clean record.

• Owned the curated analytics layer on SAP Datasphere, modelling four event streams from Parquet on ADLS Gen2 into a conformed star schema across inbound, harmonised and curated layers at 35K rows/day and 670–700 attributes per entity on a single join key, replacing manual consolidation for 10+ cross-functional consumers.
• Engineered the dbt build and test suite on the dbt-hana adapter, enforcing uniqueness, referential and accepted-value contracts in SQL plus custom tests for stage sequence and N-sigma drift, quarantining failed rows with typed exception codes.
• Designed the Power BI consumption layer over SAP HANA ODBC on hourly incremental refresh, partitioned on ingest timestamp so late-arriving corrections are never stranded, surfacing 15+ KPIs across 3 stakeholder groups, replacing 20+ hours/week of hand-built reporting.
• Orchestrated severity-tiered exception alerting through Power Automate to Teams adaptive cards with writeback acknowledgement and per-rule cooldowns, cutting exception rates 15% against an open-versus-acknowledged SLA view.

Planning ran on spreadsheet moving averages fed by a single source, and an earlier attempt to patch it had stalled. My manager handed it to me because I questioned the assumption rather than the tool. I proposed a rebuild instead of a patch, proved it on one product line in a two-week parallel run, and only then asked for full commitment. When a stale input later produced silently wrong forecasts, I flagged it the same day and rolled back. The teams that had been most sceptical trusted the system more afterwards.

• Led a 3-person rebuild of demand forecasting from Excel to XGBoost in Python on SAP BW, engineering lag, volatility, capacity-headroom and pipeline-weighted features across 500+ SKUs and 4 product lines, cutting MAPE 18% against the spreadsheet baseline.
• Devised a two-track strategy routing regular-demand SKUs to the model and intermittent-demand SKUs to a trailing annual average, validated on a temporal split with a 6-month holdout, replacing gut-based planning for procurement and capacity teams.
• Built a Python breadth-first traversal over a multi-level component hierarchy, exploding SKU forecasts into per-component monthly demand consumed directly by procurement, holding 98% on-time delivery against a 95% benchmark.
• Deployed time-series monitoring flagging demand deviations 4–6 weeks ahead across 6+ international markets, with automated alert routing into operational dashboards.
• Instituted freshness validation and output sanity checks as the pipeline's first operations after a stale input produced silently wrong forecasts, halting bad data before it reached procurement.

A short, focused internship, and where I first learned that most of the loss in a process concentrates in a handful of exception modes, provided the data is clean enough to rank them honestly.

• Structured a Pareto analysis over 9 months of operational time-series data in Python, ranking 140+ exception categories by contribution to downtime, lifting process efficiency 10% for the operations team.
• Integrated corrective-action tracking against the ranked categories across procurement, quality assurance and operations, cutting exception frequency ~90% and reprocessing rate 14% on a high-volume workflow.