Backend · Distributed systems · AI infrastructure

Nikhil Ghind

Software EngineerBay Area, CAnikhilghind29@gmail.com

I build high-throughput backend systems that hold under real load.

Backend and distributed systems — five years across trading and payments infrastructure at Interactive Brokers and NetApp. Finishing an MS at San José State, and looking for Software Engineer roles now.

I spend my time on the parts of a system that decide whether it holds up — storage engines, caching layers, and the messaging paths underneath trading and payments platforms. Five years of that at Interactive Brokers and NetApp, in Java and C++, and I am finishing an MS in Computer Science at San José State alongside it. I am looking for Software Engineer roles now.

See the work Résumé

GitHub LinkedIn

Open to SWE roles — backend, distributed systems, AI infra

BUILT AT

DAT // EXPERIENCE

Five years shipping to production

From payments and trading infrastructure at the service tier, down to storage engines and GPU datacenters.

Fall 2026
NVIDIAIncoming
Software Engineer Intern · AI and Systems Software
Joining to work on the reliability of large-scale GPU datacenter infrastructure — failure triage, telemetry correlation, and MTBF/MTBI reporting across compute clusters.
PythonBashKubernetesSlurmHPC
Scope
- Triage failures across large compute clusters
- Correlate logs and telemetry to system-level causes
- Report MTBF/MTBI to drive infrastructure work
June 2026 — August 2026
NetApp
Software Engineer Intern
Rebuilding metadata tooling for billion-file storage estates in modern C++, and cutting checkpoint cost for PyTorch models on ONTAP.
C++PyTorchSQLiteONTAP
Shipped · 03
January 2023 — June 2025
Interactive Brokers
Software Engineer II
Owned funding and Open Banking services end to end — build, review, test, release — for a platform serving 500k+ users.
JavaSpring BootKafkaKubernetesReact
Shipped · 05
July 2020 — December 2022
Interactive Brokers
Software Engineer I
Built the event-driven deposit and reconciliation pipeline behind the firm's daily cash flow, and mentored five engineers onto it.
JavaSpringKafkaPostgreSQL
Shipped · 02
- Multi-currency deposit pipeline$550M+ processed daily
- Automated reconciliation enginethousands of manual touches removed monthly
May 2019 — August 2019
Debian
Google Summer of Code — Student Developer
Wrote a Python crawler for open-source security tracking, normalising vulnerability patches across upstream repositories.
PythonOpen Source
Shipped · 01
- Open-source security crawlernormalised upstream patch data

DAT // PROFESSIONAL WORK

Built on the job

Systems shipped into production at NetApp, Interactive Brokers, and Debian.

NetApp2026
GUFI metadata indexer
8–10 hours → 45 minutes
Rebuilt the metadata engine as a multithreaded C++ indexer with a custom SQL query engine, cutting full scans across 10B+ files from most of a working day to under an hour.
C++SQLiteMultithreading
NetApp2026
Lock-free billion-file traversal
near-linear multi-core scaling
Sustained scaling on billion-file traversals with a lock-free work-stealing queue, per-thread database connections, and cache-aware memory layout in modern C++.
C++Lock-freeConcurrency
NetApp2026
Delta checkpointing on ONTAP
reduced write time and storage I/O
Accelerated model checkpointing on the ONTAP distributed storage platform by persisting only changed tensor state rather than full snapshots.
PyTorchONTAPDistributed storage
Interactive Brokers2023 — 25
Open Banking gateway
$200M+ in new business
Event-driven gateway in Java and Spring Boot integrating Plaid, fronted by React, serving 500k+ users.
JavaSpring BootPlaidReact
Interactive Brokers2023 — 25
Low-latency caching microservice
2.8s → 1.4s read latency
Caching tier on Kubernetes fronting Elasticsearch, Cassandra, and a distributed SQL store, serving ~100k requests per 10 minutes.
JavaKubernetesElasticsearchCassandra
Interactive Brokers2023 — 25
Bank-linking reliability
99.9% reliability
Hardened Kafka consumers with idempotent retries, circuit breakers, and dead-letter replay, eliminating duplicate ledger postings during outages.
KafkaJavaFault tolerance
Interactive Brokers2023 — 25
FHSA regulatory-compliance service
90% faster transaction processing
Refactored legacy workflows into a Spring Boot microservice with REST APIs and SQL-backed audit trails, shipped into production.
Spring BootRESTSQL
Interactive Brokers2023 — 25
Observability for trading infrastructure
60% lower mean time to resolve
Java log aggregation into Cassandra with Elasticsearch dashboards, across distributed trading infrastructure.
JavaCassandraElasticsearch
Interactive Brokers2020 — 22
Multi-currency deposit pipeline
$550M+ processed daily
Event-driven Java/Spring pipeline on Kafka over a distributed SQL ledger, handling 23 currencies and cutting unidentified deposits by 50%.
JavaSpringKafkaDistributed SQL
Interactive Brokers2020 — 22
Automated reconciliation engine
thousands of manual touches removed monthly
Rules-based matching engine over PostgreSQL, tuning joins, window functions, and indexes across wire and ACH deposits.
PostgreSQLSQLJava
Debian · GSoC2019
Open-source security crawler
normalised upstream patch data
Python web crawler for open-source security tracking, turning vulnerability patches across upstream repositories into structured records.
PythonCrawling

DAT // EDUCATION

Education

San José State University

Bay Area, CA

Master of Science, Computer Science

August 2025 — Expected May 2027

GPA 3.7 / 4.0

Coursework

Distributed Systems
Software Design
Machine Learning

University of Mumbai

Bachelor of Engineering, Computer Engineering

GPA 3.74 / 4.0

Coursework

Data Structures & Algorithms
Operating Systems
Probability and Statistics

SELECTED WORK

Four systems, built to hold.

Each one taken from the constraint that shaped it through to what it bought. Every project opens a page with an interactive walkthrough of how it actually works.

Orchestration2026

CloudPlay — GPU game-streaming orchestration

Four Spring Boot microservices that place live game sessions onto a pool of GPU hosts, and reclaim them the moment a session dies.

How this works

Concurrency2026

BlitzQueue — lock-free MPMC message queue

A bounded multi-producer/multi-consumer queue in C++17 that sustains 2M+ messages/sec, benchmarked against a mutex baseline.

How this works

Retrieval2026

Pythia — agentic RAG over 10k+ documents

A stateful LangGraph agent that routes, retrieves, reflects, and re-queries when its own context looks weak — cutting response latency 40%.

How this works

Storage engine2026

ClatterDB — distributed time-series database engine

A storage engine written from the bottom up in Go: LSM-tree, write-ahead log, MVCC snapshot isolation, and a hand-rolled query engine.

How this works

SYS // OPEN SOURCE

Open source

Selected repositories from github.com/nikhil-ghind. Everything here has source attached.

Distributed systems & storage

Consensus, storage engines, and coordination — mostly Go and C++.

raft-kvGo
Distributed key-value store built on the Raft consensus algorithm, with leader election and log replication.
RaftConsensusReplication
cluster-runnerRust
HPC cluster CI runner — Rust/Python job dispatcher with per-job cgroup and namespace isolation across hybrid-cloud Kubernetes.
cgroupsNamespacesHPCTerraform
hermesGo
Low-latency caching service with a three-tier LRU → Redis → Cassandra read-through/write-through cache and request coalescing.
CachingRedisCassandraLow-latency

Low-latency & systems programming

Lock-free data structures, kernel-adjacent code, and the hot path.

kernel-benchC++ / CUDA
Transformer attention micro-benchmark: naive/fused/flash GPU kernels vs a bare-metal ARM Cortex-M Q15 path, with Nsight cache-miss profiling.
CUDAFlash attentionProfilingARM
IoRingC++
Async I/O TCP server built on io_uring with lock-free structures, benchmarked against traditional syscall paths.
io_uringAsync I/OLock-free
ProcSentinelC
Linux kernel module for syscall hooking via ftrace, surfacing events through procfs and a ring buffer.
Kernel moduleftraceSyscalls

AI infrastructure & agents

RAG systems, LLM serving, observability, and ML platform plumbing.

babelPython
Multilingual RAG on Cohere Command/Embed/Rerank — hybrid dense + BM25 retrieval, cross-encoder reranking, grounded generation with citations.
CohereBM25RerankingMultilingual
diag-triagePython
LLM-integrated observability — ingests kernel/driver diagnostics into Elasticsearch and triages failures with a Claude agent over MCP.
ElasticsearchMCPObservability
trace-sleuthPython
OpenTelemetry AI-assisted debugging platform with distributed tracing across microservices via Jaeger.
OpenTelemetryJaegerTracing

Ranking, recsys & applied ML

Retrieval, ranking, and the experimentation around them.

flickpickPython
Two-tower neural retrieval + LightGBM ranking with FAISS, a Redis feature store, mSPRT A/B experimentation, and canary rollouts.
Two-towerFAISSLightGBMA/B testing
adrankPython
MMoE multi-task ranking (CTR + CVR) with realtime online training over 40+ features, scored by pCTR × pCVR × value.
MMoEMulti-taskOnline learning
spotlightPython
Two-tower dual-encoder retrieval with in-batch negatives and FAISS ANN, evaluated on Recall@K and MRR.
Dual encoderFAISSEmbeddings

Platform & applications

Services and products, including the flagship orchestration project.

cf_ai_flashtutorTypeScript
LLM-powered tutoring app built on Cloudflare Workers with edge inference.
Cloudflare WorkersLLMEdge
orbitTypeScript
Cross-platform productivity app with offline-first sync and smart notifications.
Offline-firstCross-platformSync
NidanaCSS / Python
Portal for early detection of genetic disorders using AI-assisted screening.
HealthcareML

All repositories on GitHub

DEC // TRADE-OFFS

Decisions worth defending

01 · BlitzQueue
Lock-free ring buffer
overMutex-guarded queue
Under contention the lock was the throughput ceiling, not the work. Per-slot sequence counters make the handoff safe without serialising every producer through one owner.
CostMemory ordering has to be right at every step; there is no lock to hide behind.
02 · CloudPlay
Best-fit placement
overFirst-fit placement
First fit spreads sessions across every GPU host and strands capacity on all of them. Best fit packs hosts densely, so whole machines drain and can be returned to the pool.
CostMore expensive to compute per placement, and it needs accurate live capacity.
03 · ClatterDB
LSM-tree
overB-tree
Time-series writes are almost entirely appends at the current timestamp. An LSM turns every disk write sequential and defers the tidying to compaction, off the write path.
CostReads may touch several sorted runs, and compaction spends I/O in the background.
04 · Interactive Brokers
A caching tier in front
overOptimising the existing queries
The read path crossed Elasticsearch, Cassandra and a distributed SQL store. Tuning each one had a floor; moving where the data lived did not. Latency went 2.8s to 1.4s.
CostA cache is a second source of truth, with invalidation to get right.
05 · Interactive Brokers
Idempotent retries with dead-letter replay
overAt-most-once delivery
A bank link that drops mid-transaction must not post a duplicate ledger entry. Making every consumer safe to retry took bank-linking reliability to 99.9%.
CostEvery handler needs a stable idempotency key, and the dead-letter path needs owning.
06 · NetApp · ONTAP
Delta checkpointing
overFull model snapshots
Most tensor state does not change between checkpoints, so writing all of it spends storage I/O to persist what is already on disk.
CostRestore has to reassemble from a base plus deltas rather than reading one file.

DOC // ABOUT

Background

Five years of the same instinct applied further and further down the stack.

depth over time

Product & APIs

Services & messaging

Data & caching

Storage engines

OS & kernel

Hardware & GPU

surfacemetal

Each column is a stretch of work; each filled row is a layer it touched. Hover a column for detail.

I started on money-moving systems at Interactive Brokers — deposit pipelines, an Open Banking gateway, and the caching and observability layers that kept them fast. Since then the work has moved closer to the metal: a multithreaded C++ metadata engine at NetApp, and GPU datacenter reliability at NVIDIA from Fall 2026.

Outside of that I build in the open — storage engines, lock-free data structures, agentic RAG. Most of what is on this page has a repository attached.

sys: open_to_work

SYS // CONTACT

Let's talk.

Looking for Software Engineer roles in backend, distributed systems, and AI infrastructure. Happy to talk about systems work, storage, or anything in the repositories above.

nikhilghind29@gmail.com

GitHub
@nikhil-ghind
LinkedIn
in/nikhil-ghind
Résumé
PDF

TERMINAL.EXEV2.0

System diagnostics… OK

Loading profile… 100%

whoami → nikhil ghind · backend + distributed systems · bay area, ca

uptime → 5 yrs production · $550M+/day cleared · 10B+ files indexed

status → open to software engineer roles

Awaiting handshake

Nikhil Ghind

I build high-throughput backend systems that hold under real load.

Five years shipping to production

Built on the job

GUFI metadata indexer

Lock-free billion-file traversal

Delta checkpointing on ONTAP

Open Banking gateway

Low-latency caching microservice

Bank-linking reliability

FHSA regulatory-compliance service

Observability for trading infrastructure

Multi-currency deposit pipeline

Automated reconciliation engine

Open-source security crawler

Education

Master of Science, Computer Science

University of Mumbai

Bachelor of Engineering, Computer Engineering

Four systems, built to hold.

CloudPlay — GPU game-streaming orchestration

BlitzQueue — lock-free MPMC message queue

Pythia — agentic RAG over 10k+ documents

ClatterDB — distributed time-series database engine

Open source

Distributed systems & storage

Low-latency & systems programming

AI infrastructure & agents

Ranking, recsys & applied ML

Platform & applications

Decisions worth defending

Background

Let's talk.