I'm Ashwani Kharwar, an AI Full Stack Engineer focused on backend-heavy product engineering, LLM orchestration, automation systems, and cloud-native delivery.
Portfolio
Ashwani Kharwar
Lower AI inference cost
Prompt & context optimization
Production engineering
Full-stack AI & backend systems
LLM providers orchestrated
Fallback & rate-limit handling
Cloud-native deployments
Helm, Docker & async queues
Capability Map
A practical stack for AI-native software.
The toolkit spans LLM systems, product backends, user interfaces, data stores, cloud delivery, and payments.
AI Engineering
Multi-agent pipelines, context optimization & LLM orchestration
Backend Systems
High-concurrency microservices, async worker queues & APIs
Languages & Core
Type-safe codebases, async scripts & database query languages
Data Layer
Relational schema design, vector indexing & memory caches
Cloud & DevOps
Production Kubernetes deployments, containers & CI/CD
Web & Commerce
Type-safe frontends, dynamic workflows & billing sync
Experience
Production ownership across AI, backend, and infrastructure.
Over 2 years of building autonomous multi-agent pipelines, LLM orchestrators, Kubernetes infrastructure, and automated business workflows.
- Built a multi-agent orchestration system that takes a single user prompt through planning, code generation, validation, and deployment autonomously — designing the context-handoff protocol that lets agents pass state to each other without human intervention.
- Applied context engineering across LLM pipelines by restructuring prompts, pruning context windows, and tuning token budgets, reducing AI inference costs by up to 50% while also lowering hallucinations and improving production response latency.
- Engineered a 3-provider LLM orchestrator across OpenAI, Anthropic, and Gemini with automatic failover after repeated provider failures, reducing vendor lock-in and improving runtime resilience.
- Created a restorable AI checkpoint system that records code state around agent runs, exposes rollback APIs, persists checkpoint metadata, and lets users inspect diffs or restore changes directly from chat and code-editor interfaces.
- Designed a multi-user MCP integration layer supporting secure, persistent provider connections — implementing OAuth 2.0 with PKCE and encrypted token storage, and exposing APIs for tool/resource discovery and streaming execution.
- Shipped a custom-domain system letting users attach their own domains to deployments, handling DNS verification, automated TLS via cert-manager, and Kubernetes ingress reconciliation end-to-end.
- Engineered a workflow-automation pipeline that turns raw invoice emails into synced QuickBooks transactions — chaining email ingestion, OCR extraction, LLM-based structured parsing, and Excel export — removing manual data entry from the process entirely.
- Built a direct QuickBooks integration that converts extracted invoice data into accounting transactions, removing 1 full manual re-entry step between document processing and bookkeeping.
- Built bidirectional email ingestion and dispatch scheduled via Celery Beat, supporting hourly, daily, and custom cron intervals so recurring workflows run without manual intervention.
- Led the migration from Docker Compose to Kubernetes, authoring Helm charts and configuring autoscaling and liveness/readiness probes to make deployments reproducible, scalable, and self-healing.
- Built asynchronous product-notification infrastructure using Redis-backed queues, a dedicated email worker, reusable templates, automated tests, and independently deployable Helm worker configuration.
Selected Projects
Products with real surfaces, not just demos.
A few shipped builds across AI audio, education workflows, anonymous messaging, full-stack architecture, payments, storage, and production UI.
Audinexa (AI Audio Generation Platform)
Built an AI voice-generation SaaS, supporting text-to-speech generation from both system and custom voices.
Integrated the Chatterbox TTS model through a typed OpenAPI client covering 4 core stages: prompt-based generation, voice-parameter handling, audio processing, and generated-audio storage in Cloudflare R2.
Added multi-tenant support — Google auth, organization-based access control, custom voice uploads, and scoped generation history/libraries per org
Designed a scalable full-stack architecture using Next.js, tRPC, Prisma, PostgreSQL, and FastAPI.
StudyNotion (Full Stack EdTech Platform)
Built a MERN EdTech platform covering 5 core student flows: course discovery, enrollment, lecture consumption, progress tracking, and ratings/reviews.
Developed 6 instructor workflows for course creation, section/subsection management, video uploads, pricing, publishing, and dashboard analytics.
Implemented 5 security/access controls - JWT authentication, bcrypt hashing, OTP email verification, password recovery, and role-based authorization - across 2 user roles: students and instructors.
Integrated Razorpay payment flows with dynamic coupon support, lecture consumption, and order history.
Mystery Message (Anonymous Messaging Application)
Built a web app using Next.js for creating and decoding anonymous mystery messages.
Implemented secure backend APIs with Next.js API routes and MongoDB for message storage.
Designed a clean and responsive UI with TailwindCSS for smooth user experience.
Focused on privacy and anonymity, ensuring safe one-time message sharing.
Education
Foundation and trajectory.
Have an AI product, automation system, or backend that needs to become real?
Send the rough shape. I can help turn prompts, workflows, APIs, and infrastructure into something teams can use.