Ashish Kumar Swain

• Designed a LangGraph supervisor-agent system (4 sub-agents: fetch → validate → transform → write-back) for SAP process automation — cut end-to-end task time from ~25 min to < 3 min with 100% structured JSON audit trail. • Engineered a production RAG pipeline (LangChain + FAISS + OpenAI text-embedding-3-small) over 50K+ internal documents — analyst query time dropped from 15 min to < 30 sec; adopted by 3 business teams in week 1 of release. • Shipped 8 FastAPI endpoints integrating GPT-4o and Claude (Anthropic API) with streaming and exponential-backoff retry; iterated through 3 RAGAS eval cycles and lifted answer-faithfulness score from 0.61 → 0.84. • Built Python + SQL ETL pipelines normalising 5+ mixed-format data feeds into a shared data lake — reduced downstream LLM pre- processing overhead by ~70%. • Owned prompt-engineering and model-testing cycles before each release: defined eval datasets, tracked per-domain hallucination rates, and iterated on system prompts to hit production quality gates.

• Architected a real-time Salesforce CRM ↔ marketing-automation integration (REST API + webhook pipeline) syncing lead state across 10K+ contacts — eliminated ~2 hrs/day of manual CSV reconciliation. • Replaced a brittle nightly batch export with a Python ETL layer (API pull → transform → Salesforce bulk upsert); pipeline failure rate fell from ~15% to < 1%, unblocking daily campaign operations. • Automated SparkPost + GreenArrow performance reporting with Python (openpyxl), consolidating KPIs across 4 campaigns — saved the team ~3 hrs/week and standardised metric definitions across squads.

• Built 4 Power BI dashboards (patient flow, bed occupancy, ops metrics) used in weekly executive reviews — cut report-prep time by 40% and surfaced 2 previously untracked bottlenecks in discharge workflows.

• Redesigned relational schema for a 500K-row vehicle-parts dataset and added compound indexes — SQL query runtime dropped from 8 sec → 1.2 sec, improving pipeline throughput by 6×.

• Automated weekly sales-tracking reports with Python (pandas + openpyxl), replacing manual spreadsheet assembly — cut reporting time from 4 hrs to 20 min across 12 product categories.