Languages: Go, Java, Python, C++, SQL
GenAI & Retrieval: RAG, Agent Orchestration, MCP servers & tools · Hybrid Dense/Sparse Retrieval (BM25, SPLADE), Cross-Encoder Re-ranking, Embedding Models · Retrieval Evaluation (Recall@k, offline eval) · LLM Fine-tuning & Extended Pre-training, Synthetic Data Generation · Elasticsearch, OpenSearch, MilvusDB, ANN Indexing (HNSW, IVF) · Claude Code
Streaming & Data: Apache Kafka — Kafka Connect (source/sink), exactly-once semantics, idempotent producers, consumer groups, partition & replication strategy, rebalancing, offset management, broker tuning, capacity planning · Redis Streams, Dead Letter Queues · PostgreSQL, MongoDB
Distributed Systems: Event-Driven Architecture, Microservices, Multi-Tenant Platforms, High Availability, Leader Election · API Design (gRPC, REST, GraphQL) · CPU/Memory Profiling, Latency & Throughput Optimization, Load Testing · SLO/SLA Ownership, On-Call & P1 Incident Response
Cloud & Platform: IBM Cloud, Microsoft Azure · Kubernetes (Helm, HPA), Docker · IBM Z (s390x) · Spring Boot, Resilience4j · CI/CD, Grafana, Prometheus, Jaeger
IBM India Software LabsAug 2024 – Present · Bangalore
Senior Staff Software Engineer · Tech Lead (4 engineers + 2 QEs)
IBM Bob — the RAG platform behind watsonx Code Assistant for Z: a configurable, multi-tenant embedding & retrieval service (Go, Python; IBM Cloud, x86 + IBM Z). Owned end to end — architecture & retrieval research → implementation → production rollout → operations, SLOs and on-call.
- RAG ingestion pipeline — architecture to production — Took the embedding and ingestion pipeline from ideation through implementation to production with the team: 3 Redis Streams and 4 decoupled services — consumer groups, DLQs, autoscaling, Grafana SLOs — sustaining 99.95% delivery at sub-100ms across 10K+ artifacts.
- Retrieval research & index architecture — Owned the retrieval research: ran the BM25 baseline study, then integrated SPLADE learned-sparse retrieval into a hybrid dense + sparse stack for Recall@20 +35%; evaluated index types and ANN structures (HNSW, IVF) on MilvusDB, then ran a zero-downtime migration onto hybrid Elasticsearch (x86) / OpenSearch (s390x, IBM Z).
- Retrieval performance — Cut retrieval P99 latency 50% (4s → sub-2s) and raised throughput 50% with Recall@20 and offline eval scores held flat — batched the cross-encoder re-ranking stage and fixed memory growth in the Go ONNX Runtime bindings.
- Agent orchestration & MCP server — Re-platformed retrieval behind an MCP server driven by agent orchestration — agents plan and invoke retrieval as tools instead of calling a fixed pipeline; built a novel LSP-based chunking algorithm that splits code on real syntax boundaries (Python, MongoDB, GraphQL).
- Self-serve platform adoption — Opened the platform as an in-house retrieval service any team can onboard onto and ingest into — per-team configurable ingestion (code, docs, PDFs) and isolated deployments; a Strategy-based plugin re-architecture gave 90% code reuse, 2× throughput, adopted by 4 teams at −60% integration time.
- Training data — Synthetic data pipeline yielding 50K training pairs from a 5K seed (10×) for extended pre-training and fine-tuning.
Walmart Global Tech IndiaOct 2021 – Jul 2024 · Bangalore
Software Engineer 3 · event-driven transaction platform (Java, Spring Boot, Apache Kafka, Kubernetes)
- Platform ownership — Designed and owned a fault-tolerant Java / Spring Boot event-driven stream-processing platform on Apache Kafka sustaining 1.5M transactions/day at 50 TPS and 99.99% availability — idempotent producers, exactly-once semantics, DLQ isolation, and Resilience4j circuit breakers with bulkhead isolation across gRPC and REST service boundaries.
- Kafka Connect & integration — Built and operated Kafka Connect source and sink connectors moving multiple upstream and downstream data streams through the platform — converter/transform configuration, connector-level error handling, dead-letter routing, and retry / restart semantics.
- Cluster operations & scale — Owned production Kafka operations on a 3-broker cluster (RF=3): partition and replication strategy, consumer-group tuning, rebalancing, offset management, broker tuning and capacity planning — 15+ topics, 50+ event types.
- Performance engineering — Cut p99 consumer lag 75% (2s → 0.5s) by collapsing 15+ redundant consumer queries into 3, and shrank the wire payload 67% (150 → 50 bytes) — driven by CPU/memory profiling and load testing against production traffic shapes.
- Reliability & observability — Eliminated 2 recurring SPOFs and drove SRE-led P1 incident response on centralized logging, Prometheus / Grafana SLOs and Jaeger distributed tracing — MTTR −40%.
- Scale & delivery — Ran 12+ autoscaled microservices on Azure with Kubernetes (Helm, HPA); a release-automation control plane cut deploys 7–10 days → 45s across 20+ repos.
system-design — distributed-systems design learned by building: worked design docs, runnable simulations of the trade-offs, and from-scratch implementations — a Kafka-style segmented log with sparse indexes and fenced leases, and a storage engine with MVCC and VACUUM. 57 tests, zero dependencies. site ↗ code ↗
agent-flow — agentic-AI workflow simulator: deep-research, RAG Q&A and multi-agent code-review pipelines run step by step in a streaming console with live token/latency meters. Sequential steps light up in turn while independent ones fan out. Deterministic and fully offline.
site ↗ code ↗
obsidiary — open-source publisher for Obsidian vaults: wikilinks with Obsidian's real resolution rules, transclusion, backlinks with context, and a graph. Isomorphic renderer, zero-dependency core, 199 tests. site ↗ code ↗