Machine Learning Intern

What we're building
Your ATS gets resumes in and offers out. The interview loop in between still runs on calendar invites and memory. Someone reads the resume five minutes before the call, asks whatever comes to mind, and says "seemed strong" in the debrief.
Bash runs that middle part. Before the call it researches the candidate and writes the question plan. During the call it sits beside the interviewer and suggests what to ask next. After the call it checks every claimed skill against what was actually said and writes the report in under a minute. Between rounds it carries forward what's verified and what's still open. At the end it ranks the finalists side by side, every mark cited.
Underneath that sits a lot of agent work. Rebuilding a candidate's history from a resume and scattered public sources. Generating 80 questions with ideal and weak answers for a role it has never seen. Following a live transcript and deciding when to nudge the interviewer. Grading a claim against what was said, consistently enough that two runs agree. Most of it is multi-step and tool-heavy, and it fails in ways a single LLM call never does.
The role
You'll own a scoped piece of that: building agent workflows, wiring up retrieval, and working out why a run fell over at step four.
Real work, shipped. Not shadowing someone else's.
What you'll do
Build and serve agent workflows through FastAPI services
Design and improve RAG pipelines: chunking, retrieval quality, grounding
Build and integrate MCP servers so agents can reach tools and data
Evaluate agent runs. Trace the failure, name the failure mode, fix it
Tune cost and latency across the pipeline
What you'll need
One hard requirement: Python with FastAPI. You've built and deployed APIs, not just run notebooks.
After that, the more of this you've touched the better. Nobody has all of it.
Retrieval: you've built a RAG pipeline end to end. Chunking, reranking, hybrid search
Agents: LangChain, LangGraph, MCP, tool calling, anything where the model takes more than one step
Vector DBs: LanceDB, Qdrant, or whatever you've actually run
Fundamentals: tokenization, context windows, sampling, attention, embeddings, how these models get trained
Bonus points for voice agents, multi-agent setups, or fine-tuning with LoRA or PEFT.
Who fits
You pick up new tools fast. This space changes monthly, so how quickly you learn matters more than what you already know today.
You follow a problem through to done. You flag blockers early, and nobody has to ask you twice.
Worth knowing
Agent work is mostly debugging. Something works on the third try, then breaks the moment it hits a real candidate's data.
We're small, so there's no ticket queue waiting for you. You'll get a problem and a deadline, and you're expected to say when you're stuck.
Do well and you'll get a pre-placement offer, before the internship ends.
Apply
Send your resume and a link to something you've built to scribble@bash.ai. Projects beat GPA here.
Bash AI is an equal opportunity employer.

Work arrangement
No