Gaurav Kumar

• Architected a multi-agent predictive-maintenance platform: a LangGraph supervisor routing to six specialised agents discovered through a capability registry, with intent classification and a semantic cache ahead of the orchestrator so repeated queries bypass planning and inference entirely.
• Designed a layered retrieval strategy that routes each query to vector, graph, full-text or hybrid search and unifies candidates through a reranker — replacing indiscriminate fan-out across every backend.
• Built dual failure-prediction subsystems: calibrated ML models for failure category, time-to-failure and downtime, alongside an LLM failure-mode classifier with three-tier retrieval fallback, so degraded dependencies reduce answer quality instead of raising errors.
• Established the security and network architecture for managed agent infrastructure: least-privilege IAM, private-subnet deployment over VPC endpoints, and a systematic access-verification harness covering every platform component across control and data planes. Used IAM policy simulation to evaluate destructive permissions without executing them, converting an ambiguous access failure into a reviewable list of policy changes for the security team.
• Delivered a self-hosted voice pipeline (ASR → LLM → TTS) behind a FastAPI gateway with edge audio normalisation, gRPC to model services and a streaming WebSocket path, keeping operational audio inside the network.
• Owned feature architecture and data governance for a supply-planning build-feasibility model, gating feature inclusion on measured population rates with an explicit recorded rationale per attribute.
• Led decomposition into independently deployable services with no shared business logic, paired with module-federated micro-frontends.

• Built a document intelligence platform for extracting structured data from unstructured documents — the problem space that led directly into the retrieval and ingestion work above.
• Developed web-scraping agents with self-healing behaviour, recovering automatically when target sites changed structure instead of failing the run. Brittleness against upstream change is the dominant failure mode in automation, and designing for it is the same instinct behind the tiered retrieval fallbacks in my later agent work.
• Four years of platform engineering in enterprise automation and orchestration — the domain that most directly anticipates agentic systems.

Client-facing systems architecture and delivery across enterprise engagements.

• Agentic RAG backend with hybrid pgvector and full-text retrieval fused via Reciprocal Rank Fusion, chosen over weighted score blending because cosine distance and text-rank scores share no scale and no fixed weighting survives an embedding-model change.
• Page-aware Docling ingestion with boilerplate stripping and table-aware chunking, hardened against a 167-page tender document containing repeated page furniture, self-contradictory clauses and commercially significant tables.
• Fixed a silent data-loss defect where position-independent content hashing caused 166 of 167 page headers to be discarded by an upsert conflict rule with no error raised.
• 42-question evaluation set spanning multi-hop, table, contradictory and unanswerable categories; 169 tests running without network, database or LLM dependencies, plus CI with lint, coverage across two Python versions, pgvector integration and dependency auditing.

• LLM-assisted financial planning that keeps projection arithmetic deterministic in code and confines the model to interpretation and explanation.

Manipal University