SDE intern (Frontend)

Brainwave: AI Software Engineering Intern

Team: Brainwave (AI Product Engineering)  |  Location: Bangalore  |  Type: Internship (Full-time, on-site, 5 days WFO)  | 

About MediBuddy

MediBuddy is India's leading digital healthcare platform. We bring doctor consultations, diagnostics, medicine delivery, and health checkups into one connected experience for people across the country.

Every line of code Brainwave ships shapes how someone gets an answer, gets guided through a surgery journey or a vaccination appointment, or gets unblocked faster. That responsibility is why our culture runs on one belief: #wehealbycode.

About the Role

Brainwave builds and owns Mia, MediBuddy's Intelligent Assistant. Today that means customer query resolution. Soon it means personalized experiences across surgery assistance, vaccination journeys, and other moments where people need fast, accurate, individually relevant guidance.

This isn't a role to learn what an LLM is on the job. Come in having already built something real with one, a chatbot, an agent, a RAG pipeline, and knowing why it worked or didn't. What you'll actually develop here is judgment. That means engineering context so the model is reliably right instead of confidently wrong, debugging a multi-step agent pipeline down to its real cause, and turning messy domain knowledge into structure a system can use.

We care less about seniority than about trade-off engineering. Can you reason about the right call given accuracy, latency, and what a real user needs? A mentor will guide you, but don't stop at "what the mentor said." If you see a better call, make the case for it. Do that well and you'll get real ownership, and a real seat at the table.

What You'll Do

• Work across the stack: the backend services powering the CS chatbot and adjacent internal tools.

• Get hands-on with our agent architecture. Understand the graph, the nodes, and the state before you touch any of it.

• Debug real, live issues: a wrong tool call, a broken handoff between agent nodes, a retrieval step that pulled stale context. Trace it to the real cause, whether that's bad chunking, a prompt regression, or a schema mismatch, instead of papering over it with a bigger prompt.

• Sit regularly with stakeholders across the business (PMs, CS agents, ops folks who know the process) to find gaps in current SOPs and turn their tacit knowledge into context the system can actually use.

• Build MCP servers on top of our existing internal systems, whether that's booking, order status, catalog, or whatever the workflow needs, so the agent gets standardized access to live data instead of another brittle glue script.

• Work on context engineering: deciding what the model should see, when, and in what form, so it gives grounded answers instead of confident guesses. That includes building and reading evals, not just eyeballing outputs and calling it good.

• Reason through real tradeoffs: latency vs. accuracy, longer context vs. tighter retrieval, a bigger model vs. a better prompt. Be able to defend the call you made.

• Use AI-assisted coding tools as a normal part of how you work. Know when to trust them and when not to.

• Take ownership of specific parts of the assistant as you prove you understand how it behaves and why.

Requirements

1. A real understanding of how LLMs work under the hood: context windows, tokens, why models hallucinate and how you actually mitigate it, why retrieval fails as often as it succeeds (stale embeddings, bad chunking, wrong top-k), and what makes a structured output reliable versus a schema violation waiting to happen. "I've used ChatGPT" doesn't count.

2. At least one substantial AI/LLM project you've built or meaningfully contributed to: a chatbot, an agent, a RAG pipeline, something with real API calls and real data moving through it. Not a UI wrapper around someone else's endpoint.

3. Strong debugging instincts. Comfortable reading logs and traces, tracing a bad output back through a multi-step pipeline, forming a hypothesis before guessing.

4. Some notion of how to evaluate an LLM system beyond "it looked right in the demo": a test set, a regression check, some way to know if a prompt change actually helped.

5. Genuine, hands-on comfort with AI-assisted coding tools. You already have a working workflow, not casual exposure.

6. High attention to detail, a fail-fast mindset (no assumptions, make mistakes, recover fast, learn), and real curiosity about how these systems behave, not just how to make them demo well.

7. Bonus: you've personally hit and diagnosed an LLM failure mode (hallucination, context loss, a bad tool call) and understand why it happened.

Tech Stack

Orchestration & agents

• LangGraph, our agent runtime

• LangChain, used where it fits alongside LangGraph

Model & tool access

• LLM APIs (Anthropic / OpenAI or similar),MCP

Evals & observability

• Eval and tracing tooling to measure output quality and catch regressions before they ship

Backend & infra

• FastAPI (Python) / Next.js / Express or Nest.js 

• AWS

Prompt and context engineering is core to the job, not a tool on a list. See What You'll Do and Requirements above.

Interview Process

Round 1 - Assignment (Take-Home) 

Round 2 - Python + Context Engineering (Live, No AI Assistance)

Round 3 - Open Round (Manager & Culture Fit)

Why Join Us

• Real ownership of a live, in-production AI product that's actively growing. Not a sandbox project.

• Direct mentorship from the engineers building the system, with real context on why it's built this way and where it's headed. A mentor sets a floor, not a ceiling: question decisions, bring a better idea when you have one, push the bar up alongside them.

• Frequent, direct access to CS agents and real user conversations. Rare for an intern.

• A domain where a better chatbot answer means a patient or customer gets help faster.