Vedant Andhale.
I design and ship reliable AI systems — retrieval, agents, evaluation, and the backend infrastructure around them.
I care about the parts that make model capability usable in production: data contracts, APIs, failure handling, observability, performance, and human review.
Selected work
Engineering proof, with the decisions left in.
FreightSense: making shipment recommendations inspectable
A FastAPI decision-support prototype combining deterministic shipping calculations, LLM recommendations and a human override history.
EvidenceDeterministic rules + LLM comparison, explicit disagreement handling, API fallbacks and append-only human overrides.
Crop Cure: a grape-leaf classifier inside WhatsApp
A research prototype connecting a PyTorch image classifier to a multilingual WhatsApp workflow with FastAPI.
EvidenceWhatsApp → FastAPI → PyTorch inference, multilingual response flow and an explicit low-confidence fallback.
loadfile: one call to load tabular data
A small Python package that loads local and cloud files into pandas through a consistent API, replacing repeated file-loading code.
EvidenceOne API across local and cloud tabular data with fsspec backends, multiple file formats and ZIP member handling.
Notebook
Notes from building, debugging and measuring systems.
Returning to open source: collaboration first, code second
Why I am returning to open source in 2026, what I learned from EddieHub, and how I want to contribute differently as a working engineer.
First-Year Computer Science Roadmap: What I Wish I Knew
A practical first-year CS guide to DSA, AI, projects, placements, referrals, GATE, internships, and building strong software engineering fundamentals.
Keep a static-site build easy to explain
The build contract behind this Hugo portfolio: pinned inputs, explicit environments and generated HTML.
Before caching a result, define when it is wrong
Cache keys, stale data and invalidation are part of the result contract, not just performance settings.