Dhanush Chandra ShekarDhanush C.

Managing OpenAI, Anthropic, and Gemini separately created fragmented infra with inconsistent latency and no fault isolation.

Built a unified FastAPI gateway with semantic caching, intelligent routing, and circuit breakers behind a single endpoint.

98% reduction in p95 latency with zero-downtime failover and full Prometheus observability.

Teams using Slack, Notion, and GitHub had no shared memory — context was siloed and data conflicts went undetected.

Engineered a 12-agent LangGraph system backed by Neo4j to manage organizational state and flag conflicting transactions in real time.

Cross-platform state resolution across 3 integrated environments with automatic conflict detection.

Technical hiring relied on manual code reviews, creating bottlenecks and inconsistent candidate evaluation at scale.

Built an AI-driven platform with VS Code integration, Docker-sandboxed code execution, and automated evaluation pipelines via Claude AI.

Full recruitment workflow from submission to structured scoring, fully automated end to end.

Standard vector search over SEC filings missed multi-hop relationships between entities, reducing answer quality for complex queries.

Combined Neo4j graph traversal with vector embeddings to enable entity-aware, multi-hop Q&A retrieval over financial documents.

Outperformed vector-only baseline on context extraction across a 46-query benchmark suite.

Off-the-shelf NMT models are opaque — hard to study attention or customize training without abstraction layers in the way.

Built a complete encoder-decoder Transformer from scratch in PyTorch with custom attention, positional encoding, and training pipeline.

Fully functional sequence-to-sequence translation model trained end-to-end with zero external NMT dependencies.

Retail demand forecasts degraded silently over time — no retraining triggers, no drift monitoring, no deployment automation.

Built a full MLOps pipeline with XGBoost/LightGBM ensemble, automated retraining, drift detection, and CI/CD via GitHub Actions.

MAE of 0.2575 with zero-touch deployment and continuous monitoring in production.

Mental health discourse on Reddit is nuanced and class-imbalanced — generic classifiers fail to categorize it reliably.

Fine-tuned DistilBERT with NLPAug data augmentation to handle severe class imbalance and improve cross-category robustness.

72.5% F1 accuracy on a highly imbalanced multi-class mental health classification task.

Raw transactional data had no dimensional structure — BI reporting was slow and analytical queries were hard to maintain.

Designed a star schema and built a production ETL pipeline with PostgreSQL and SQLAlchemy, surfaced via Power BI dashboards.

Analytics-ready data layer enabling self-serve BI with structured dimensional queries across the full dataset.

Marketing decisions were based on noisy experiment data — no rigorous statistical framework to validate conversion lifts.

Built an automated A/B testing framework with scipy statistical testing, ETL reporting cycles, and Tableau dashboards.

12% conversion rate lift validated through statistically significant testing with automated reporting.