I am Divyansh Rawat, an AI & Backend Engineer pursuing Electronics & Communication Engineering at IIITDM Kurnool.
I specialize in architecting distributed backend services in Go and Rust, alongside autonomous multi-agent systems in Python and TypeScript. My core technical focus centers on distributed systems, agentic AI, and high-performance deep learning inference.
I focus on full-lifecycle systems engineering from architecture to production operations. I engineered and deployed the cloud infrastructure and reservation engine powering Quadis Hotels, handling live bookings on AWS in production.
At Blipai, where I serve as a Growth & AI Engineering Intern, I built an ultra-low-latency speech-to-text pipeline powered by Whisper on Groq LPU inference (~700ms end-to-end) across macOS, Windows, and mobile platforms.
Beyond production systems, I author technical breakdowns on software architecture and actively practice competitive programming on LeetCode and Codeforces.
Growth & AI Engineering Intern ·
BlipaiApr 2026 — Present
Built the real-time speech-to-text pipeline behind a product that ships on macOS, Windows, iOS and Android. Whisper served on Groq LPU inference, ~700ms end to end in production.
Tuned decoding parameters and added a custom-vocabulary prompt layer, fixing technical and domain-specific terms the base model kept mangling.
Ran the feedback loop with the product community: triaged feature requests and bug reports, ranked by impact, shipped what users actually needed.
Debugged platform-specific failures across macOS, Windows, and Android (installers, OS permissions, audio devices, broken configs).
Education
Indian Institute of Information Technology, Design & Manufacturing, Kurnool — B.TechAug 2024 — 2028
Electronics & Communication Engineering
Featured Work
Quadis Hotels
Live in production
A booking platform for a Delhi NCR hotel group: rooms, rates, real payments, admin
back-office. I designed it, built it, deployed it and ran it, alone. It is live today on
infrastructure I stood up and later handed to the client.
The whole stack is mine: a React 18 + TypeScript front end on Vite, a
Node/Express + TypeScript booking API, and a PostgreSQL schema
covering properties, room types, rate plans, holds and bookings.
Production ran on my own AWS account first (S3 + CloudFront + Elastic Beanstalk + RDS).
I then moved all of it onto the client's account: a hardened EC2 box in ap-south-1
with an Elastic IP, nginx, systemd and Postgres. Scripted build/push deploys, a rehearsed DNS cutover.
Razorpay checkout with webhook signature verification, PDF invoicing and WhatsApp confirmations. Secrets from SSM Parameter Store, auth paths fail closed.
Hardened the box: TLS & security headers at nginx, rate limiting, Helmet, Zod validation, systemd unit locked down with ProtectSystem=strict.
Quadis Assist is an LLM concierge on Groq with deterministic fallback.
Memory tracer bounds validation, ES rollover lookback protection, and trace anonymizer atomic flush pipeline.
Problem: Memory configuration accepted zero or negative MaxTraces limits, enabling unbounded in-memory allocations; in jaeger-anonymizer, reaching --max-spans-count caused an immediate exit before flushing the UI trace capture to disk; es-rollover lookback with unit count ≤ 0 wiped all indices from the read alias; API v2 client crashed due to missing gogo codec registration.
Engineering: Implemented strict positive bounds validation with unit test coverage; restructured anonymizer execution lifecycle to guarantee complete atomic writes upon span limit exhaustion; registered the gogo protobuf codec; pinned regexp query flags for the Elasticsearch search engine.
Cloud-native cluster automation, read-only CI formatting enforcement, and node readiness controller fixes.
Problem: Prow CI's verify-gofmt modified files in-place during validation rather than running read-only, masking formatting regressions across Kubernetes repos; Node Readiness Controller Helm chart hardcoded probe port 8081 ignoring custom bind configurations; leader election Role & RoleBinding failed outside default namespace; RuleReconciler returned RequeueAfter with error, violating controller-runtime contracts; unvalidated taint values caused node patch failures.
Engineering: Hardened verify-gofmt to execute strictly read-only and fail fast on unformatted code; corrected Helm chart RBAC namespace templates; aligned controller reconcile return contracts; introduced CEL validation for taint value formats.
OpenAPI v1 schema compliance (KEP-0004 §2), IEEE 754 cursor pagination precision, RFC 8594 deprecation, and training runtime test suites.
Problem: Kubeflow Model Hub v1 APIs silently fell back to default sort when receiving undeclared orderBy parameters instead of returning 400 Bad Request required by KEP-0004 §2; Hugging Face sync intervals of 0s triggered server panics; float-valued cursor pagination lost precision, inducing infinite query loops; preview pagination stalled on duplicate model names; alpha catalog endpoints omitted RFC 8594 Deprecation/Sunset headers.
Engineering: Engineered strict order_by.go OpenAPI schema validator rejecting invalid fields with 400; preserved IEEE 754 double precision in base64 cursor tokens; added RFC 8594 headers; developed comprehensive test suites for TrainingRuntime & ClusterTrainingRuntime validation webhooks.
Prefix cache indexer verification test harness and multi-node LLM serving orchestration for distributed inference gateways.
Problem: Unit test suite for the prefix cache indexer was disabled in upstream CI due to timing flakiness in distributed gateway mock states; lack of standardized multi-node inference documentation caused orchestration divergence.
Engineering: Isolated mock KV-cache index states, eliminated race conditions, and restored CI coverage for distributed prefix caching and prompt reuse across multi-tenant GPU workers; authored official guides for Kubernetes external metrics and multi-node orchestration.
Embodied AI telemetry, safety interlock audits, ARM64 hardware compatibility, and model export pipelines.
Problem: Critical systems & security flaws identified: joint velocity clamps silently disabled after any prior joint violation; ZeroMQ server bound to 0.0.0.0 with an unauthenticated kill endpoint; unauthenticated cascade revocation endpoint; model verification pipelines traced native weights instead of export artifacts; Jetson Orin ARM64 install failures due to unpinned dependencies.
Engineering: Resolved Jetson Orin ARM64 dependencies; expanded embodiment kinematic schema for drones/quadcopters; audited and diagnosed 20+ safety interlock and distributed training vulnerability issues directly with upstream maintainers.
AWS session caching, atomic role locking, and distributed multi-cloud credential isolation.
Problem: High-frequency cloud automation workflows triggered upstream STS rate-limits and token expiration races during concurrent role assumptions; lack of atomic credential locks caused session corruption across worker threads.
Engineering: Diagnosed and specified atomic IAM role session locking and deterministic token cache invalidation to prevent rate starvation and eliminate duplicate STS requests.
