You bring the
hard problem.
I ship the fix.
Backend and AI/ML engineer, recent CSE grad. I build the systems underneath the product — APIs, data pipelines, LLM infra — and I ship with the same ownership on day one as on day ninety.
Proof, not adjectives.
Numbers from shipped work — not a skills bar chart.
How you can hire me.
Startup-shaped roles. High ownership across the stack.
Backend / API Engineer
- API design & data pipelines
- Caching & query optimization
- Service architecture
- Production deploys
AI/ML Systems
- Applied ML pipelines
- Model evaluation & benchmarking
- LLM security testing
- Containerized inference APIs
Early-Stage Generalist
- Full-stack when the team is small
- Own a feature end-to-end
- Fast iteration, no hand-holding
- Comfortable with ambiguity
Security-Minded Eng
- Prompt injection testing
- LLM security testing
- Adversarial robustness
- Responsible disclosure
Selected work.
Three internships, one independent security disclosure.
Backend for a multi-tenant compliance platform (ISCC)
Software Development Intern at Carboledger, an early-stage startup (<10 employees). Converted requirement cards into production-ready backend features, applying design patterns and modular architecture to build extensible, maintainable components. Owned identification and resolution of critical bugs impacting workflow accuracy, edge cases, and system reliability across cross-version testing and regression cycles.
RAG pipelines for AI-driven backend features
Software Engineer Intern at Infinoid Technologies. Built Node.js backend services powering AI features in production. Implemented RAG pipelines with LangChain + PyTorch, using PostgreSQL + pgvector for embedding storage, indexing, and similarity search — plus document chunking and metadata-based filtering to sharpen retrieval accuracy for LLM Q&A. Applied security guardrails and auth controls to AI/backend endpoints, deployed on GCP.
API latency reduction across 12 REST endpoints
Web Development Intern at Fast Conversion eSolutions (remote). Reduced API latency via Redis session caching and lazy-loaded non-critical Express middleware. Replaced blocking synchronous operations with async equivalents to cut event-loop stalls. Refactored 12 REST endpoints to eliminate N+1 queries and add connection pooling.
Independent LLM security research — CodeVerdict
Identified and responsibly disclosed a critical system prompt leakage vulnerability in CodeVerdict, a Claude-based code review platform, via adversarial input structuring against LLM context-boundary trust assumptions. Root cause, reproduction steps, and remediation assessed by the founder as immediately actionable; publicly recognized for report quality and depth.
Things I ship on my own.
Personal projects — built, not just started.
RedisLink — URL shortener
Node.js/Express service with a Redis–MongoDB two-tier caching layer, connection pooling, and async I/O for concurrent request handling. Helmet.js, CORS, XSS-safe input sanitization, idempotent operations. Cache-first strategy with 24h TTL, error middleware, health checks, graceful shutdown, automated retry logic.
Interview Transcript Summarizer
CLI tool that converts raw interview transcripts into structured recruiter summaries with topic coverage, inferred role/level, and concise candidate takeaways. Generates validated summaries for sample transcripts and saves reusable JSON outputs.
vector-search-api
Production-ready semantic search API that ingests documents or URLs, generates embeddings with OpenAI, stores them in PostgreSQL + pgvector, and powers namespace-aware natural language retrieval with auth, rate limiting, and caching.
CDM-based fake image detection
Led a research team implementing a CVIP 2021 CDM-based autoencoder for fake colorized image detection — parallel encoder processing R–G, G–B, R–B channel difference maps with cascaded scale blocks and transfer learning from regeneration to detection. HTER-evaluated across ImageNet and Oxford Building.
Toddler autism prediction system
Containerized behavioral classification API — Soft Voting Ensemble (Random Forest + Naive Bayes) — achieving 0.95 F1 with real-time probability scoring for early diagnostic support.
Who you'd be working with.
I'm Ayush — a backend and AI/ML engineer, recent B.Tech CSE graduate (IIIT Vadodara, 2022–2026). I gravitate toward the parts of a product that other people avoid: the data pipeline nobody wants to own, the agent that needs guardrails before it ships, the endpoint that's quietly O(n²).
Outside of internships, I co-founded the Droid Robotics Club (scaled to 50+ members) and worked as a Technical Member at MLSA, conducting practical workshops on RESTful backend development for 20+ peers — API architecture, request handling, and cloud service integration. I write technical breakdowns on LinkedIn and Hashnode — RAG-to-agentic retrieval, LLM threat modeling, coding-agent design — grounded in primary documentation, not summaries of summaries.
I hold an Atlassian Agile certification and completed the Stanford / DeepLearning.AI ML Specialization at 98%.
I take the problem, not the task list. I flag risk early and ship — you shouldn't have to chase me for status.
What working with me looks like.
I own the outcome
Give me the problem, not a spec. I find the path, flag risk early, and ship.
Correctness over speed-theater
The right call made once beats the fast call unwound three sprints later.
Decisions in writing
Trade-offs documented as I make them — so context isn't locked in my head.
Security by default
Threat-model AI and data-facing surfaces before they ship, not after an incident.
Have a backend or AI problem that needs to ship?
No deck, no fluff — just the problem, and whether I'm the right person to own it. I read every email myself.