I build AI systems that engineers can argue with.
I work on the parts of AI systems that have to be correct: retrieval pipelines, execution paths, state machines, and what happens when they break.
One shipped product, one internship, and several systems I can walk through end-to-end — including where they failed and what I changed.
Akhil Pavuluri
Systems Engineer
I got pulled toward systems work because the interesting failures were never in the demo.
They showed up in the boundary layers: unclear semantics, weak contracts, silent drift, and execution paths nobody could inspect.
That is the work I like most: turning ambiguous model behavior into explicit system behavior.
Primary domain
Applied AI in real products — retrieval, execution paths, state management.
What I've built
RAG pipelines, order state machines, graph-based debugging tools, governed SQL analytics.
Bias
A boring stack that works over clever code that breaks under pressure.
Contrarian stance
I do not trust fluent output unless the stack can expose the path that produced it.
AI / ML Intern — Tech Bharat AI
Sep 2025 – Jan 2026LLM systems on Vertex AI: ingestion → chunking → retrieval → ranking → grounded generation, orchestrated with LangGraph for internal tools and assistants.
- Pipeline shape: document ingest, embedding stores, re-ranking, and constrained prompts so answers cite retrieved spans — not free paraphrase.
- Hard tradeoff in practice: wider retrieval improved F1 but blew P95 latency; addressed with tiered retrieval and smaller reranker batches.
- Shipped internal assistants and tools tested against real documents — inconsistent formatting, varied structures, noisy scans.
Solo developer — Printish
2024End-to-end printing platform: auth, payments, file handling, real-time order tracking, and admin operations — TypeScript and React.
- Customers upload files, pay, and track orders to delivery. Admins manage fulfillment and catch stuck jobs before users notice.
- Hardest part: order state when couriers don't update on time — status copy that stays honest when you don't actually know where the package is.
B.Tech — Artificial Intelligence & Data Science
KCG College of Technology · 2022 – 2026
1st Place — National Hackathon
CARE College
YOLO-based Smart Traffic Management System optimized for real-time inference.
1st Place — Hack Bell Cybercratz 2.0
Hackathon
Full-Stack E-Learning Platform architected and deployed under strict time constraints.
Determinism over probability when accountability matters.
Structured representations over raw text when semantics must survive scale.
Explicit failure modeling over silent assumptions.
Observability before optimization.
Systems that remain understandable age better than systems that merely look clever.
Let's build something that still makes sense under pressure.
If the problem involves semantics, system behavior, reliability, or AI that has to survive production, that's the right conversation.
Download resume