Backend Engineer Intern
Bengaluru
About TRADL AI:
TRADL AI is building India's first AI trading co-pilot for 100M+ traders.
Today, we convert natural language into verified, code-backed stock intelligence - no hallucinations, no black boxes. Tomorrow, we're building the full stack: real-time chart intelligence, autonomous research agents, and portfolio-aware execution. We're turning what institutional desks spend millions on into a conversation anyone can have. We're live, growing fast, and the hardest technical problems are still ahead of us.
Why This Role Exists
TRADL's AI capabilities are expanding faster than a single team can ship. We need someone who can operate across the stack, from fine-tuning models and building agentic workflows to wiring those capabilities into the product. This isn't a shadow-and-observe internship. You'll own real features, push to production, and work directly with the engineering team on problems that don't have textbook answers.
Specialisation Tracks
Track A: Claude Code & AI Tooling. You understand agentic workflows, build with skills and sub-agents, configure memory via CLAUDE.md, manage context windows deliberately, wire up MCP servers, and use hooks to intercept and extend Claude Code's behaviour.
Track B: Model Fine-Tuning. You've fine-tuned at least one model end-to-end, from data preparation through training to evaluation. You understand the difference between full fine-tuning, LoRA, and QLoRA, know when to apply SFT vs alignment methods like DPO, and can interpret a training run beyond just watching the loss go down.
Track C: Both (PREFERRED) You can set up a LoRA fine-tuning run and use Claude Code agents to automate the evaluation pipeline around it. You've wired hooks, managed context with CLAUDE.md skills, and spun up MCP integrations, while also understanding what's happening inside the model.
What We're Looking For
Python proficiency. Clean, modular code beyond notebooks. Scripts, modules, CLI tools.
Claude Code working knowledge. You've used it for real tasks, not just one-shot prompts. Familiarity with CLAUDE.md, skills, and basic tool use is expected.
Agentic workflow understanding. You know what a sub-agent is, how context windows get consumed, and why that matters when chaining multi-step tasks.
Hands-on with Hugging Face. Transformers, datasets, peft, or trl. You've loaded, trained, or evaluated a model with it.
Data wrangling. Pandas and NumPy at a working level. You can build a clean instruction dataset from raw data.
Git fluency. Branches, PRs, meaningful commit messages.
Bonus Points
Multi-agent pipelines. Orchestrator delegating to specialised sub-agents with clean context passing and graceful error recovery.
Hooks. Pre/post-tool hooks to intercept agent actions: logging, filtering outputs, enforcing constraints, or triggering side-effects.
MCP. Integrated one or more MCP servers connecting Claude Code to external tools, databases, or APIs in a live agentic session.
Advanced fine-tuning. QLoRA or LoRA beyond toy examples. Preference data + DPO. Any GRPO, RLVR, or SFT-to-alignment pipeline experiment.