Systems & backend engineer — I build low-level infrastructure from scratch: consensus protocols, LLM inference servers, and distributed tracing platforms. Currently pursuing a B.Tech in Computer Science at IIEST Shibpur (CGPA 8.67, expected 2027) and interning as a Software Engineering Intern at Accenture.
I like problems where correctness and performance both matter — the kind you validate with fault injection and benchmark under load.
- Distributed systems — Raft consensus, quorum replication, linearizable reads
- AI infrastructure — transformer inference, PagedAttention KV-caches, CUDA/Triton kernels
- Observability — OpenTelemetry ingestion, tail-based sampling, high-throughput storage
- Backend & cloud — Go, Python, Java services on OCI/AWS/GCP with real SLOs
RaftKV — Distributed Key-Value Store · Go · Raft · gRPC · BoltDB · Docker
Implemented the Raft consensus protocol from scratch: leader election, log replication, snapshotting, and a durable BoltDB write-ahead log. Lease-based linearizable reads tolerating ⌊(N−1)/2⌋ node failures, validated across 50+ fault scenarios with a Jepsen-style checker — zero safety violations. 51.2K ops/sec, 1.63 ms p99 on a 5-node cluster (YCSB-C).
InferEngine — LLM Inference Server · Python · PyTorch · CUDA · Triton
OpenAI-compatible transformer inference server with continuous batching, streaming generation, and a PagedAttention-style KV-cache. Scheduler dynamically merges prefill/decode phases; custom Triton fused-QKV kernel. Within 1.18% of vLLM throughput on Qwen2.5-7B (A100) — 975.8 vs 987.4 tok/s at comparable p99 latency.
TraceFlow — Observability & Tracing Platform · Go · OpenTelemetry · gRPC · ClickHouse
OpenTelemetry-compatible trace ingestion backend with W3C context propagation and batched ClickHouse persistence. Tail-based sampling keeps 100% of error/slow traces while dropping 93.76% of healthy traffic. 128.8K spans/sec, 99.12% delivery, and 9 ms p99 trace-by-ID lookups over a 30M-span dataset.
I contribute upstream to the infrastructure projects I work with — 34+ repositories across production orgs including @NVIDIA, @nytimes, and @earendil-works. My activity skews toward real engineering work: ~49% commits, ~27% code review, ~20% pull requests.
Notable contributions:
- NVIDIA/NemoClaw — secure agent runtime on NVIDIA OpenShift
- cv/mcs and cv/go-inflect — plus 30+ other upstream repos
- AI-infra I build against and contribute to: vLLM, TensorRT-LLM, Chroma
- Distributed_LRU_Cache — distributed LRU cache ·
Python - tsdb_project — a time-series database ·
Python - code-review-assistant — automated code review tooling ·
Python - AI-interview-simulator — voice-interactive interview coaching (basis for my IEEE paper) ·
JavaScript - Code_Summarizer_AI — LLM-powered code summarization ·
JavaScript - Cloud_data_lake — cloud data-lake pipeline
Languages — Python · Go · C++ · Java · SQL · TypeScript · JavaScript Backend — FastAPI · Node.js · Express · Django · gRPC · REST Systems & AI Infra — CUDA · Triton · PostgreSQL · Redis · Prometheus · Linux Cloud & Data — AWS · Azure · GCP · OCI · PySpark DevOps — Docker · Kubernetes · GitHub Actions · PyTest · Jest
- Codeforces Master · CodeChef 4★ · LeetCode Guardian
- 2,500+ problems solved across platforms
Voice-Interactive Interview Coaching with Hybrid RAG and STAR-Based Feedback S. Ghosh, S. Maity (Supervisor: Dr. Tamal Pal, IIEST Shibpur) — under review at IEEE Transactions on Artificial Intelligence.
- 📧 ghoshsouvik9433@gmail.com
- 💻 GitHub · 💼 LinkedIn · 🧑💻 Codeforces · 🍳 CodeChef · 🟨 LeetCode