#01 · RAG
Retrieval & applied AI
Hybrid retrieval, reranking, deterministic orchestration and LLM-as-judge evaluation — RAG and multi-agent systems held to production numbers, not demo vibes.
Mudit
Garg
Delhi
India
[ 00 ] · Open to opportunities
Building backends that hold load
Backend engineer who ships production systems end to end — high-performance APIs, retrieval pipelines and multi-agent AI, on infrastructure that stays up.
Backend Engineer & Applied AI — 2026
ScrollBackends, agents and infrastructure built for real traffic.
[ 01 / 04 ]·About
From code to production
Backend engineer who designs and ships production-grade systems end to end. Most recently I owned the FastAPI and PostgreSQL backend behind a production RAG assistant answering 10,000+ questions a month across 20+ countries.
My work sits where backend fundamentals meet applied AI: high-performance APIs, hybrid retrieval, deterministic orchestration around LLMs, and the evaluation that proves any of it works. Day to day that means Python, Go and TypeScript on PostgreSQL, Redis and Kafka, shipped with Docker, Kubernetes and Terraform.
I care how a thing is engineered rather than whether it ran once in a demo — circuit breakers, retry budgets, resumable pipelines, profiling the hot path. Outside work I build systems for the sake of understanding them: an nginx-style reverse proxy from raw net/http, and a sandboxed platform where an autonomous agent attacks deployed code.
[ Remote ] · Delhi → Dubai
0KM
Delhi to Dubai: 2,205 km as the crow flies.
- From
- Delhi, IN
- 28.61°N · 77.21°E
- Shipped to
- Dubai, UAE
- 25.20°N · 55.27°E
Based in Delhi, building production backends for a Dubai property company — fully remote, four months, 98 migrations and 1,100+ tests later.
#02 · API
Backend systems
FastAPI, NestJS and Go services on PostgreSQL, Redis and Kafka — schema design, caching, rate limiting, auth and pipelines that resume where they failed.
#03 · INFRA
Infra & reliability
Docker, Kubernetes, gVisor and Terraform; Prometheus, Grafana and OpenTelemetry — deploys that are boring and outages that stay small.
Skills
Languages
01PythonGoTypeScriptJavaScriptJavaSQLC++
Backend & APIs
02FastAPINestJSNode.jsExpressREST APIsSSEEvent-drivenMicroservicesOAuth / JWT
Data & Messaging
03PostgreSQLpgvectorRedisMongoDBKafkaRedpandaBullMQDrizzle ORMAlembic
AI & LLM
04RAGLangGraphLangChainMCPMulti-agentCohere RerankLLM-as-judgeLangfusePyTorch
Cloud & DevOps
05DockerKubernetesHelmgVisorTerraformAWSGitHub ActionsVaultLinux
Observability
06OpenTelemetryPrometheusGrafanaProfilingLoad testingFuzzing-raceTracingAlerting
Areas of expertise
- Backend architecture & APIs/
- Retrieval-augmented generation/
- Multi-agent orchestration/
- Distributed data & messaging/
- Reliability at scale/
- Performance engineering/
- Cloud infrastructure & DevOps
Professional experience
-
#1 · Apr. 2026 - Jul. 2026
Full Stack Developer Intern
OffplanWise (Anav Consulting) · Remote, Dubai
Owned the FastAPI + PostgreSQL backend (~36K LOC, 81 endpoints, 1,100+ tests) of a production RAG chatbot at offplanwise.com. Designed dual-KB hybrid retrieval — pgvector HNSW and Postgres full-text legs fused with RRF and a Cohere rerank — lifting recall from 62% to 86%. Built the turn orchestrator, three fault-tolerant ingestion pipelines, and a 117× faster listings ETL.
- Hybrid retrieval · 62% → 86% recall
- ~84% faithfulness, ~90% citation accuracy
- ETL clean stage 78 min → 40 s
- −45% retrieval latency, −35% LLM spend
- OAuth, rotating JWTs, 4-layer guardrails
-
#2 · During B.Tech
Core Team Member
Microsoft Tech Community · Amity
Mentored 100+ juniors through data structures, backend fundamentals and their first projects, and ran coding contests for 300+ participants.
- Mentored 100+ juniors
- Contests for 300+ participants
Education & certifications
-
Degree
B.Tech, Computer Science & Engineering
Amity University, Noida
-
Certification
Fundamentals of Deep Learning
NVIDIA Deep Learning Institute
-
Certification
Cybersecurity Professional
Google
Key numbers
01
Recall lifted from 62% to 86% on a 131-question eval set with dual-KB hybrid retrieval
02
Listings ETL clean stage cut from 78 minutes to 40 seconds — 117× — by batching ~50K round-trips
03
Retrieval latency down 45% and LLM spend down ~35% with caching and a 77% prompt-cache hit rate
04
Retry budget held load amplification to 1.12× (vs 3.0×) through a full backend outage in Prism
05
1,000 concurrent requests collapsed into a single backend fetch by RFC 9111 request coalescing
06
Real exploits landed on OWASP Juice Shop by an autonomous agent inside a gVisor sandbox
[ 02 / 04 ]·Work
Selected projects
A selection of platforms, infrastructure and AI systems — designed, built and measured end to end.
Prism — HTTP Reverse Proxy in Go
An nginx-style reverse proxy (~25K LOC) written without httputil.ReverseProxy: streaming, WebSockets, four load balancers, circuit breakers and an RFC 9111 cache. 68.8K req/s proxied, 99K on cache hits.
Solar-Wind-LSTM — Storm Early Warning
An attention-based LSTM that forecasts geomagnetic storms from solar-wind telemetry six hours before they arrive.
[ 03 / 04 ]·Record
Ranked, not random.
330+ problems solved · 1,955 contest rating · top 3.37%
Let's stay in touch
Open to backend and applied-AI roles, and to interesting systems problems. The inbox is the fastest way to reach me.
- Emailmuditgarg.tech@gmail.com
- Webmuditgarg.xyz
- BaseDelhi, India