KAVYASRI JADALA kavya2693

Senior Data Scientist | Demand Forecasting · Supply Planning · Inventory & Logistics Optimization · Applied AI

8.5+ years across retail, supply chain, grocery/quick-commerce and healthcare · Dubai, UAE (Golden Visa) · M.S. Purdue

Portfolio · LinkedIn · kavyasri.2693@gmail.com

`MATCH (me) RETURN who`

CREATE (me:DataScientist {
  name:      "Kavyasri Jadala",
  role:      "Senior Data Scientist",
  focus:     ["Demand Forecasting", "Supply Planning",
              "Inventory & Logistics Optimization", "Applied AI"],
  also:      ["Knowledge Graphs", "Graph-RAG", "Agentic AI"],
  domains:   ["Retail", "Supply Chain", "Grocery / Quick-commerce", "Healthcare"],
  based_in:  "Dubai, UAE  ·  Golden Visa",
  education: "M.S., Purdue University"
})
RETURN me.rule;

// ▸ "Score the model on the decision it drives, and report where it loses."

Featured work

Each repo: problem-first README and results including what didn't win; most also carry an architecture section, tests and CI.

Forecasting & demand

Repo	Problem	Headline result
intermittent-demand-forecasting	Which forecasting method to run per SKU when nearly half the catalogue sells nothing in most weeks	Weekly WAPE 58.1% → 42.5% vs seasonal naive across 1,500 SKUs
order-forecast-inventory	Customer × SKU order forecasts for a B2B distributor, judged by the stock they let the warehouse hold	Holding cost at 98% fill −46.8% vs ETS
retail-demand-forecasting	Store × SKU replenishment forecasts, with an ablation on whether a product/store graph earns its place	WMAPE 0.2273 → 0.1838 (19% lower) vs seasonal naive; graph lift only +0.4%
peak-capacity-planner	Riders per grocery hub and shift through Ramadan and a 4x mega-sale, checked order by order in a replay	On time 76.9% → 98.9% of accepted orders, 17.6% less spent on riders

Inventory & supply planning

Repo	Problem	Headline result
ddmrp-buffers	How much stock a lubricants distributor releases with DDMRP buffers, at the service level customers get today	Days of inventory 97.6 → 52.1 (−47%) at matched 98.46% fill
agentic-supply-planning	Forecasts turned into budgeted, approved purchase orders by a LangGraph supervisor and four agents	88% touchless PO rate on the retail range
supply-chain-twin	Test replenishment, promotion and fleet changes on a SimPy twin of a retail chain before any store sees them	−16.6% total cost on the tuned quarter; the same plan fails with 20% of the fleet out

Routing & last mile

Repo	Problem	Headline result
fleet-routing-cvrp	Next-day routes for a 300-truck freight fleet with OR-Tools, and which lever actually cuts misrouted shipments	Peak-week road km −22.1% vs sweep baseline
last-mile-dispatch	30-minute grocery dispatch across 26 dark-store hubs, judged by a hub-level A/B test	Fleet cost per delivered order −1.04 AED (−10.6%)

Pricing & commercial

Repo	Problem	Headline result
markdown-optimizer	How deep to mark down each slow-moving grocery SKU, from its own elasticity instead of a blanket 30% off	Promotional margin +37,750 ± 8,252 vs the 30% rule
order-outcome-margin	Which marketplace orders will lose money, and when to switch off cash on delivery	COD margin rule +22,258 (+0.72%) where blocking all COD loses 16.6%
hcp-graph-targeting	Which doctors a launch sales team should call when prescription deciles stop separating converters	Converters in a 5% call plan: 126 (decile list) → 212 (graph features + embeddings)
genai-recommender	Next content, channel and compliant message for 20,000 prescribers	BERT4Rec NDCG@10 0.303 loses to a last-item Markov chain at 0.316

Agents, LLMs & data platforms

Repo	Problem	Headline result
llm-cost-router	Send each LLM request to the cheapest self-hosted model likely to get it right, and escalate only on a failed check	61% cost saved vs all-large; completion 82.5% vs 85.0%
agentic-pipeline-builder	Messy source exports in; a typed, tested dbt project, a quarantine and an Airflow DAG out	Column types 54/54 (100%) vs 44/54 for DuckDB auto-detect
novacart-mlops	Churn, demand forecasting and a Graph-RAG assistant served as one MLOps stack	Churn ROC-AUC 0.795; logistic regression beat gradient boosting (0.772)
predictive-maintenance	Which pumps, compressors and heat exchangers to service this week on a six-job crew budget	Cost per year 5.27M → 2.02M vs run-to-failure

Knowledge graphs & governance

Repo	Problem	Headline result
customer360-ontology-graph	A customer-360 graph that rejects malformed entities at ingestion and answers questions from the graph, with citations	SHACL gate rejects the injected fault
ontoloop	Adverse drug event narratives into a knowledge graph, repaired until a SHACL validator passes	3B model conformance 0% → 85% with the loop; accuracy does not improve
mdm-entity-resolution	The same doctors across CRM, offline and web data resolved into golden records	Pairwise precision ~0.93, recall ~0.84
commercial-metadata-catalog	Glossary, catalog and lineage that trace a marketing slide back to its source content and approval	Demonstration; no headline metric
blastradius	Column-level lineage that reaches the application code a column change will break	100% precision, 90% recall (27 / 0 / 3) on labelled references

`cat /skills`

Forecasting, ML & optimization

LLMs & agents

Knowledge graphs

Data platforms & MLOps

`./stats.sh`

`cat /certs`

`tail -f /connect`

"A forecast is only as good as the stock, routes and orders it changes. I measure that part."

Navigation Menu