AI Research Engineer Intern DeepMechanix

Location: Remote, India-based · US HQ: San Jose, CA Duration: 3 months, May–August 2026 Stipend: ₹85,000 – ₹1,25,000/month, set on demonstrated depth Return offer eligible: Yes — full-time role + RSUs for strong performers 

We are building superintelligence for the design and engineering of physical infrastructure, the hardware that civilization runs on. We are starting with static equipment: pressure vessels, heat exchangers, reactors, and related pressure-containing equipment that the global chemical, energy, and pharmaceutical industries depend on.

Engineering workflows in this domain are largely unchanged since the 1990s. Engineers transcribe datasheets by hand, perform ASME code calculations in decades-old desktop applications, and produce deliverables through manual review cycles that can take weeks. We are building the software layer that replaces the manual parts of this workflow while preserving the rigorous compliance and traceability that industrial customers require. Our first product combines custom retrieval over ASME standards with automation of engineering software. We are pre-seed with revenue, targeting a seed round in late 2026.

Compensation: Monthly stipend of ₹75,000–₹1,25,000/month for a 3-month engagement. The top band is reserved for candidates demonstrating both research-grade depth and production engineering quality.

Return offer and equity exposure: Structured as a pipeline to a full-time AI/ML Engineer role. Strong performers receive a return offer with increased compensation and an RSU grant tied to the company's exit outcome.

Mentorship and authorship: Direct technical collaboration with the founding team. Novel research contributions, building on the ASME PVP 2026 framework, are eligible for co-authorship on papers or patent filings.

Letter of recommendation: A detailed letter of recommendation useful for applications to top US and European research programs (e.g., Stanford, CMU, MIT, Berkeley, ETH Zürich, EPFL).

Resume signal: Experience with a US-incorporated deep-tech AI startup, a novel regulated industrial domain, founder-level technical collaboration, published external work, and equity exposure.

You will own the retrieval, reasoning, and evaluation systems that transform foundation models into reliable engineering tools.

In this domain, hallucinations are design liabilities, not cosmetic bugs. Your core challenge is engineering the harness that ensures frontier LLMs remain reliably correct within regulated, standards-driven contexts.

Active workstreams during the internship:

• Retrieval pipeline evaluation: The core RAG pipeline is live and in production. A significant part of this role is designing the evaluations that drive decisions about whether to keep, replace, or redesign individual components, running those evaluations, and defending the resulting recommendations with data.

• Agentic calculation pipeline: Driving automated engineering calculations through an agentic system that is deliberately built with pinned fallback data for cases where RAG retrieval is insufficient. Ongoing work includes formalizing when to trust retrieval versus fallback, and improving that boundary.

• Evaluation methodology: Building on the evaluation framework from the ASME PVP 2026 paper. Substantial contributions during the internship are eligible for co-authorship on follow-on papers.

• Context engineering and structured output: Improving how the system handles edge cases in ASME standard parsing, table extraction, and structured output fidelity where general-purpose LLM behavior is insufficient.

Our mandatory operating requirement is fluency with agentic coding tools. We ship daily using Claude Code, Cursor, Codex, and adjacent agent-first environments. Onboarding establishes these patterns, but strong candidates have already internalized and leveraged them to compress implementation cycles.

Working hours: Primarily async, with a 30-minute daily synchronous standup call during the 8–11 AM IST window.

Core stack:

• Production-grade Python: async patterns, type annotations, proper error handling, testing discipline.
• Transformer architecture from first principles: attention mechanisms, positional encoding, tokenization, training dynamics. Not API familiarity, architectural understanding.
• End-to-end RAG, context engineering, or multi-agent system experience: At least one project where you made and defended engineering system design choices, for example in a RAG project
• AI layer: Prompt engineering, context control, structured instruction design, retrieval systems (embeddings, vector databases, RAG-based workflows using manual pipelines and frameworks such as LangChain, LangGraph etc.)
• LLM-assisted development: Code generation, refactoring, debugging, code review, and architectural reasoning using Cursor and Claude Code
• Python backend: Core Python, OOP, unit testing, pip, virtualenv, async patterns, FastAPI
• APIs and models: Claude, GPT, Gemini, Huggingface, Landing AI, Reducto
• Data layer: PostgreSQL, Azure Blob Storage
• Tooling: CLI, Git fundamentals (branching, merging, pull requests), GitHub Actions CI
• Document parsing: Azure Document Intelligence, pymupdf, VLM-based extraction, Landing AI, Hybrid Parser

Familiarity and strong ability to use agentic coding tools can make up for unfamiliarity with specific items in the stack above.

Enrolled in masters or phd, exceptional btech students will also be considered.

The bar is high and the definition of exceptional is intentionally broad. Strong candidates typically fit one of these profiles, but the profile matters less than the proof of exceptional talent.

• Builder with a shipped AI system Experience with a production AI system (e.g., RAG pipeline, fine-tuned model, custom evaluation framework) that demonstrates real engineering design decisions; repositories must show technical rationale, not just output.

• Research-oriented engineer Published paper or substantive arXiv pre-print in NLP, information retrieval, or ML systems, and the proven ability to cleanly implement new techniques in Python within days. First-author work is a strong signal.

• Competitors You have a proven track record of execution under pressure, evidenced by hackathon wins or competitive programming results.

• Academic Achievers CGPA 9.0+ in CS from a top-tier institution (IIT/BITS/NIT) with mastery of core systems subjects, paired with independent building evidence.

Attach a PDF CV. Include in the body:

1. A link to your GitHub profile
2. Three to five specific examples of your strongest technical work, each with a direct link; competition leaderboards, paper URLs, repository URLs, deployed product URLs
3. A single paragraph on the one technical project you are most proud of, and why
4. A short note (3–5 sentences) on the agentic coding tools you use today, how you use them, and one specific example where they materially changed how you approached a problem

Selection process:

1. 15-30 minute screening call
2. 90-minute technical interview
3. 30-minute founder conversation on fit and working style