Website:
auricai.in
Job details:
About Auric AI
We're building an AI system for defense and intelligence that reads thousands of messy, multilingual documents (field reports, intercepted communications, news articles, company filings) and automatically discovers hidden connections across them that no human analyst could find by reading alone. The stakes are real. The problems are unsolved. And we intend to build the best system in the world for it.
The hardest part isn't RAG. It's entity resolution: determining that five different mentions across five documents are the same real-world person, even when one is a full name, another is an abbreviation, another is a Hindi transliteration, another is a vague description, and one has no name at all. Get this wrong and the entire system breaks. Most teams underestimate this problem. We don't.
The Role
You own the AI pipeline end to end: extraction from unstructured text, entity resolution, knowledge graph construction, hybrid retrieval, and proactive discovery of threats nobody asked about. A separate team handles infrastructure/MLOps. Your focus is purely on the AI.
We hire very few people and hold an extremely high bar. Everyone on the team can independently design and build systems that have no standard playbook. We don't hire for potential. We hire people who are already exceptional and give them problems worth their time. If you're looking for a role where the problems are defined and the architecture is handed to you, this isn't it.
The Hard Problems
- The same person appears across documents as a full name, an abbreviation, a Hindi transliteration, an honorific, and a vague description with no name at all. Your system decides what merges and what doesn't, knowing that a wrong merge is far worse than a missed one.
- Documents are mixed Hindi-English, OCR'd from handwriting, and follow no standard format. You design extraction that turns this into structured entities and relationships with hallucination detection.
- The knowledge graph grows with every document, self-corrects bad merges, and maintains full provenance. Tens of thousands of entities, full source traceability.
- Users ask pattern questions spanning multiple hops: "Is there evidence of foreign organizations using business fronts to contact military personnel?" You design retrieval that combines graph traversal with semantic search, including agentic multi-step retrieval.
- The most valuable capability: proactively scanning the graph for similar patterns across unrelated entities, detecting anomalies, surfacing emerging threats. Nobody asked for these discoveries. The system finds them on its own. This isn't RAG. Nothing like this exists yet. You'll design it.
You Might be a fit if
- You have strong AI/ML fundamentals and can reason from first principles about problems you haven't seen before
- Given an ambiguous, open-ended problem, you decompose it into layers, identify tradeoffs, and design a system. Not look for a tutorial.
- You understand how LLMs, embeddings, and retrieval systems work at a deep level. Not just the APIs.
- You've built AI systems that had to work on real, messy data. Not just clean benchmarks.
- You think about failure modes before being asked. You know that a system working 95% of the time can still be useless if the 5% failures are catastrophic.
- You can design novel systems when there's no paper to follow and no library to import
- Specific experience with entity resolution, knowledge graphs, NLP, or multilingual systems is a strong plus but not a prerequisite. We care more about how you think than what you've already built.
Our Team
A small, exceptional group of researchers and engineers building at the frontier of AI for national security. The strongest technical minds we could assemble here because the problem matters and the bar is the highest they’ve encountered.
We’re setting the global standard for AI-driven intelligence analysis by building smarter architectures than anyone else in this space.
Click on Apply to know more.