// identity loaded

CHITRANSH SAXENA

SR. STAFF ENGINEER · GENAI × DISTRIBUTED SYSTEMS

I build production retrieval systems that don't fall over — RAG, embedding pipelines, exactly-once streaming at enterprise scale. The architecture and the taste are mine; the machine holds the boilerplate. A poet who ships RAG.

7+ yrs shipped3+ yrs genai1.5M txns/day+35% recall@20
Download résumé
assistant · online

summarize Chitransh in one line

Senior staff engineer shipping production RAG & exactly-once streaming — architecture and taste by hand, boilerplate by machine.

01 identity record

A poet who ships RAG.// 01

I hold the architecture and the taste; the machine holds the boilerplate. Production GenAI on distributed infrastructure — and a soft spot for things that bend through the air.

··leetcode solved ··github commits ··public repos
··site visits
handle@chitransh-saxena
roleSr. Staff SWE — IBM
focusRAG · embeddings · streaming
baseBangalore, IN · UTC+5:30
exp7+ yrs · 3+ GenAI · 1.5+ Claude Code
toolingClaude Code daily, 1.5+ yrs
status◉ open to hard problems
  • signal RAG · +35% recall@20
  • scale 1.5M txns/day · 99.99%
  • off-hours FPV drones · guitar · photography
  • on grass trivela free-kicks · juggling

02 experience.ledger — open a row

Three latent shelves.// 02

01 IBM Software Labs Sr. Staff SWE 2024 — now BLR RAG · +35% recall@20

IBM Bob RAG Team — watsonx Code Assistant for Z, configurable SaaS embedding & RAG platform · Go codebase

  • Before this was a funded initiative, built the seed for it myself — a GitHub scraper + chunker + embedder feeding a small RAG chatbot over our own repos; it spread across teams organically for months before being scaled into this platform.SELF-STARTED
  • Architected & led (4 devs · 2 QEs) an end-to-end embedding + RAG pipeline ingesting 8K+ code pairs, 50+ docs and 250+ runtime examples.8K+ PAIRS
  • Integrated SPLADE learned-sparse retrieval after a deep BM25 study, with sub-second hybrid inference.+35% RECALL@20
  • Prototyped vector search on MilvusDB (HNSW/IVF, quantization), then ported to a hybrid Elasticsearch (x86) + OpenSearch (s390x, IBM Z) stack via a zero-downtime migration.<900MS
  • Profiled Go inference services for memory issues in ONNX Runtime bindings and batch-optimized the reranking stage.P99 4S → SUB-2S
  • Redis Streams backbone with a novel LSP-based chunking algorithm over 10K+ artifacts via custom MCP tools, with a React status dashboard and Slack alerting for pipeline health.SUB-100MS · 99.95%
  • Re-architected a monolith; adopted by 4+ teams. ★ Star of the Month in 2 months.90% REUSE
  • Opened it up as an in-house retrieval service any team can onboard onto and ingest into — configurable ingestion (docs, code, PDFs), per-team deployment.SELF-SERVE
  • Synthetic-data pipeline, TDD-validated, for LLM extended pre-training — profiled and re-threaded it for an 80% throughput gain.50K FROM 5K · 10× · +80% THROUGHPUT
  • Mentored and directed an intern in extending the synthetic-data tooling into a sampling/inspection tool — small-scale test runs, sample-data review — later adopted by product and upper management for POC evaluation.MENTORSHIP
  • Shipping it in Go (Golang) — actively leveling up the language on the job.

GoRAGSPLADEMilvusDBElasticsearchOpenSearchs390xONNX RuntimeRedis StreamsMCPReact

Fault-tolerant event streaming, hiring-platform leadership, and retail-scale delivery

  • Designed & owned a fault-tolerant, event-driven stream platform sustaining 1.5M txns/day @ 50 TPS on 3-broker Kafka (RF=3) — idempotent producers, exactly-once semantics, DLQ isolation.99.99% AVAILABLE
  • Led design & delivery (3 engineers, 2 QAs) of Walmart's rehire-eligibility system — caveat rules, past-history checks, work-remark checks — across UI, backend and DB, shipping to 100K+ applicants a year.100K+ APPLICANTS/YR
  • Led a full front-end + back-end revamp of Walmart's internal Hiring Centre (React, Java, Azure) — better UX, wider device support, and real infra cost savings.FULL REVAMP
  • Designed the HLD for a Kafka Connect ETL pipeline (Strategy-pattern connectors across multiple databases) streaming every job-posting event to Phenom for syndication across Indeed, LinkedIn and other boards in every country Walmart hires in — owned capacity planning and SLAs end to end.MULTI-COUNTRY ETL
  • Cut p99 consumer lag by collapsing 15+ redundant queries into 3; compressed the wire format 67% (150 → 50 bytes).P99 2S → 0.5S
  • Architected an extensible event-processing framework (Strategy/Factory/Decorator, Resilience4j circuit breakers, bulkhead isolation, backoff retries).50+ TYPES · 15+ TOPICS
  • Ran 12+ autoscaled microservices on K8s; built a MERN release-automation control plane across 20+ repos.7–10D → 45S
  • Eliminated 2 recurring SPOFs; drove SRE-led P1 incident response (logging + Jaeger tracing).MTTR −40%

KafkaKafka ConnectExactly-onceKubernetesAzureResilience4jJaegerReactMERN

Core C++ engine, on-prem infra ops, and monolith decomposition at data scale

  • Built and optimized the core C++ data engine generating client-facing projections — CPU cache-level optimizations and a migration from plaintext parsing to Protocol Buffers cut serialization overhead across huge data files.TEXT → PROTOBUF
  • Designed a NATS-based work-distribution pattern — one primary dispatches sub-jobs as messages, a C++ worker pool executes them in parallel, results fan back in and aggregate — alongside custom allocators and hot-path string↔integer conversion.FORK/JOIN · MQ
  • Implemented a TL-planned re-architecture off a single C++ monolith into horizontally scalable services, coordinated by leader election — core computation stayed C++ for raw performance, a new Java/Spring Boot layer took over orchestration.THROUGHPUT +60%
  • Replaced legacy Autosys scheduling with Apache Airflow DAGs — cut licensing spend, raised CPU utilization, and gave the team far more scheduling flexibility and control.−$100K/YR
  • With everything on-prem, owned the "cloud" layer by hand — nginx reverse-proxying and load balancing, Redis caching, microservice deployment & scaling, DB management at scale, mounted disks, endianness.ON-PREM · FULL-STACK OPS
  • Re-engineered latency-critical C++ pipelines across 15K+ LOC — profiling, leak elimination, hot-path restructuring; owned production on-call, release validation and triage.−55% RUNTIME
  • Co-designed and built a full test pyramid — smoke, integration, performance, load — with 4 other engineers.4-DEV TEST BUILD

C++Protocol BuffersNATS MQJava/Spring BootApache AirflowMicroservicesLeader electionOn-call

edu — B.Tech Information Technology · SRM Institute · 2015–2019 · 88%

03 stack.matrix — click a cluster to expand its tools ›

The competency lattice.// 03

00GenAI / MLRAGEmbedding modelsSemantic & hybrid searchBM25 / SPLADERe-rankingMultimodal ingestSynthetic dataFine-tuningPrompt engAgentic codingMCP11
01Retrieval / ANNFaissHNSWIVFQuantizationANN indexingHybrid retrieval06
02DatabasesPostgreSQLMS SQL ServerMongoDBElasticsearchOpenSearchMilvusDB06
03Streaming / DistributedApache KafkaKafka ConnectRedis StreamsExactly-onceEvent-drivenDLQsProtocol BuffersMicroservicesLeader electionCircuit breakersHASRE & observability12
04FrameworksSpring BootSpringResilience4jgRPCREST APIsNode / ExpressReact07
05LanguagesJavaPythonC++GoSQLNoSQL06
06InfrastructureKubernetesDockerHelmAzureApache AirflowGrafanaPrometheusJaeger08
07Personal / Self-hostedGitHub SCMGitHub ActionsGitHub IssuesGitHub ProjectsCloudflare DNSCloudflare WorkersCloudflare Zero TrustWAF / SSLOracle Cloud · OCIDockerTerraformWrangler12

04 how i build

I ship with Claude Code in the loop.// 04

For 1.5+ years it's been my default tooling — agentic coding, custom MCP servers, rapid POCs across enterprise & personal work. More shots on goal.

“I hold the architecture and the taste; the machine holds the boilerplate.”
  • POCs in hours, not weeks
  • Custom MCP tools wired into real data
  • Tests-first, even for throwaways

↳ this site was designed & shipped with Claude Code, served from Cloudflare's edge.

// retrieval_pipeline
RAG retrieval pipeline Five stages, left to right: ingest, chunk via LSP, embed, index, retrieve. ingest chunk·LSP embed index retrieve // idle — press RUN_QUERY to trace a retrieval
chitransh@edge — claude
~/portfolio claude "render my résumé as a calm console"
✓ Astro · Lenis momentum scroll · GSAP reveal
✓ light / dark themes · magnetic controls · ⌘K palette
✓ Inter · Fraunces · JetBrains Mono
→ shipped to the edge
~/portfolio ▍

05 projects.index · 5 live

Sub-constellations.// 05

01● LIVE

system-design

Distributed-systems design, learned by building. Each problem is worked end to end — requirements, NFRs per tier, capacity maths, and several complete architectures rather than one, each optimal under a different dominant constraint. On top sits a concepts layer: the ~15 generative ideas that recur across every problem, so an unfamiliar system becomes something you re-derive instead of recall. The trade-offs are then made playable — crank an ad firehose and watch OTP p99 blow past its 1s budget on a shared cluster, then switch topology and watch it stay flat. First build: a multi-tenant MQ platform at 550K msg/s with a 158× cost-per-message spread across tiers.

vanilla JS/CSS · zero deps · zero build · runtime-rendered markdown

live ↗code ↗

02● LIVE

obsidiary

An open-source publisher for Obsidian vaults — wikilinks with Obsidian's real resolution rules, transclusion, backlinks with context, a graph, and .canvas files rendered as pages. Your notes stay in their own repo and the engine is just a dependency, so upgrading never touches your writing. The renderer is isomorphic, so the same code also reads any public vault straight in the browser with no build at all.

TypeScript · Astro · remark · zero-dep core · 199 tests

live ↗code ↗my notes ↗

03● LIVE

agent-flow

An agentic-AI workflow simulator — pick an agent pipeline (deep research, RAG Q&A, multi-agent code review, support triage) and press play. Sequential steps light up in turn while independent ones fan out and run in parallel, an agent console streams the trace, and an "under the hood" panel names the real algorithm & libraries behind each step. Fully mocked & deterministic — no live model calls, runs offline.

vanilla JS/CSS · zero deps · zero build

live ↗code ↗

04● LIVE

trivela

A free-kick game that makes the Magnus effect impossible to misunderstand — curl it round the wall into the top bins. Vanilla web component, embeddable; play it live or read the source.

JavaScript · canvas · physics

live ↗code ↗

05● LIVE

deadlock-dinner

Interactive, kid-friendly visualization of the Dining Philosophers problem — locks, mutexes, semaphores & deadlock come alive, with real Java & Go code.

JavaScript · concurrency viz

live ↗code ↗

06 arcade — a tiny game, built into the page

Don't clip the hurdles.// 06

07 library — recent reads · pick a book

The open stacks.// 07

// the notes behind the shelf live at notes/reading ↗ — an Obsidian vault, published with obsidiary

08 open channel

Got a hard problem in
GenAI or distributed systems?// 08

→GitHubgithub.com/Chitransh-Saxena ↗
→LinkedInlinkedin.com/in/chitransh22 ↗
→LeetCodeleetcode.com/u/chitransh22 ↗
▦Résumépreview · download pdf ↓