Senior Applied AI & NLP Engineer

RedGraphs, Inc.; Delhi; Hybrid; INR 3,000,000 - 5,000,000; Full time

Job summary

RedGraphs, Inc. is hiring a senior AI engineer to build production AI systems that transform regulatory and corporate documents into structured relationships.

You will own major pieces of the extraction stack, handle evidence-based retrieval, and ensure dependable production behavior with a hybrid approach blending retrieval, models, and rules. Based in Delhi NCR, you will collaborate with the India engineering team, require overlap with US Eastern time, and work as the role shifts from remote

Qualifications

• Senior level ownership in ML/NLP or related software systems.
• 4+ years of relevant professional experience preferred.
• Strong production Python engineering skills.
• Hands-on NLP/ML/LLM, information extraction, or document intelligence experience.
• Experience moving models from notebooks to production workflows.
• Understanding of model evaluation, precision/recall, regression testing.
• Solid software fundamentals: testing, Git, Linux, APIs.
• Ability to diagnose unfamiliar behavior and communicate findings.

Responsibilities

• Improve relationship discovery and extraction from corporate documents.
• Build and improve evidence retrieval, candidate generation, ranking, and extraction.
• Enhance entity identification, resolution, and relationship attribution.
• Handle linguistic phenomena including negation, uncertainty, and multi-entity statements.
• Extract and validate financial attributes, percentages, dates, and metadata.
• Design systems blending LLMs, NLP, retrieval, classifiers, and deterministic validation.
• Create evaluation datasets and regression suites from real false positives/negatives.
• Diagnose model failures at the mechanism level and maintain existing behavior.

Skills

Senior ML ownership

4+ years experience

Production Python

NLP systems

Information retrieval

LLM experience

Model evaluation

Software fundamentals

Diagnostics & debugging

Research background valued

Tools

Git

Linux

APIs

CI/CD tests

Job description

Understand the evidence. Build production AI systems that read ambiguous corporate language and return evidence-backed, correctly attributed economic relationships.

Location: Delhi NCR, India

Employment type: Full time

Team: Engineering

Compensation: INR 30L to 50L annual CTC, with some flexibility for candidates with unusually strong technical scope and experience.

Working Model

Workplace: Currently remote, transitioning to hybrid.

RedGraphs is building its India engineering team in Delhi NCR and plans to establish a collaborative development office there. You must be based in, or willing to relocate to, Delhi NCR, and should expect regular in-person collaboration once the office is operational.

The role also requires meaningful and consistent working overlap with the United States Eastern time zone.

About RedGraphs

RedGraphs builds economic intelligence infrastructure that makes business relationships visible, measurable, and actionable.

We transform evidence buried in regulatory filings, financial disclosures, and other primary source corporate documents into structured, time-aware relationship data.

Our platform spans three connected layers. RedGraphs Assist governs how economic relationships are captured and verified from source evidence. Business Relationships Analytics, jointly developed with S&P Global Market Intelligence and commercially licensed and delivered through S&P Global Market Intelligence, provides structured supplier and customer relationship data with historical depth. RedGraphs Vista operates above the relationship graph to analyze multi-tier dependencies, economic exposure, concentration, and changing network structure.

Accuracy matters at every layer. A technically valid result can still be wrong if the evidence, entity attribution, relationship direction, economic magnitude, or timing is wrong. Our systems therefore combine machine intelligence, deterministic software, temporal data infrastructure, graph analytics, and governed human verification.

We are a small engineering team. Senior engineers are expected to understand the business meaning of the systems they build and to own important problems from investigation through production verification.

About the Role

RedGraphs processes corporate disclosures in which economically important relationships are expressed indirectly, inconsistently, and with substantial linguistic ambiguity.

Your job is to make those documents understandable to machines without losing the evidence or the meaning behind the result.

You will work on systems that identify relevant evidence, resolve entities, determine relationship type and direction, interpret financial and temporal language, and decide when the available evidence does or does not support a structured relationship.

The system is deliberately hybrid. Depending on the problem, the correct approach may involve retrieval, language models, smaller classifiers, fine-tuned models, deterministic rules, structured generation, or a combination of them.

This is not a prompt engineering role and it is not a pure research position. Research quality thinking is valuable, but success is measured by dependable production behavior on difficult real-world documents.

What Success Looks Like

During the first 90 days, you should be able to:

• Understand the major RedGraphs document-to-relationship workflows.
• Reproduce and explain at least one meaningful false-positive or false-negative class, from the evidence through to the final output.
• Establish the correct evaluation baseline before changing the mechanism.
• Ship a contained extraction or qualification improvement with appropriate regression protection.
• Demonstrate measurable improvement without silently degrading established behavior.
• Begin owning an important retrieval, extraction, qualification, resolution, or evaluation mechanism.

Over time, you should become the technical owner of a major part of the RedGraphs extraction intelligence stack, helping set its architecture, evaluation standards, and production direction.

How We Work

Evidence before assertion: Important technical conclusions should be supported by code, data, tests, or directly observed behavior.

Correctness without artificial slowness: We move quickly, but not by hiding uncertainty or bypassing validation.

Ownership beyond implementation: Engineers follow changes through testing, deployment, verification, and operational outcomes.

Independent thinking: Senior engineers are expected to challenge assumptions and recommend better approaches when the evidence supports them.

Written first: Investigations, decisions, architecture, and operational knowledge should remain reviewable by other people.

Historical integrity matters: Improving current behavior must not silently change what historical data means, or what information was knowable at a given point in time.

This Role Is Not

This is not primarily:

• chatbot development
• prompt writing
• generic generative AI application integration
• research without production ownership
• model experimentation without rigorous evaluation

We are looking for someone who wants to make AI behavior measurable, explainable, evidence bound, and dependable.

Responsibilities

• Improve relationship discovery and extraction from regulatory filings, financial disclosures, and corporate documents.
• Build and improve evidence retrieval, candidate generation, ranking, classification, and extraction mechanisms.
• Improve entity identification, entity resolution, relationship attribution, and relationship direction.
• Handle difficult linguistic phenomena including negation, qualification, uncertain language, historical statements, indirect references, and multi-entity sentences.
• Extract and validate financial attributes, percentages, monetary values, dates, and other relationship metadata.
• Design systems combining LLMs, traditional NLP and ML, retrieval, classifiers, and deterministic validation, and determine when a probabilistic model is appropriate versus when deterministic software is the safer approach.
• Develop reviewed evaluation datasets and regression suites based on actual false positives and false negatives.
• Diagnose model failures at the mechanism level rather than treating aggregate accuracy as sufficient.
• Improve precision and recall while protecting already correct behavior.
• Design structured model outputs that can be validated before publication and traced back to the document and text that support them.
• Improve model inference reliability, latency, throughput, and cost where relevant.
• Work with platform engineers to move AI capabilities into idempotent, observable, production workflows.
• Document assumptions, failure modes, evaluation methodology, and architectural decisions.

Required qualifications

• Strong evidence of senior level technical ownership in applied machine learning, NLP, AI engineering, or closely related software systems.
• Typically 4+ years of relevant professional experience, or equivalent evidence of senior-level technical scope and ownership. We care more about demonstrated technical depth and ownership than years of experience.
• Strong production Python engineering.
• Hands-on experience building NLP, machine learning, LLM, information extraction, search, or document intelligence systems.
• Experience taking model behavior beyond notebooks or demonstrations into a tested production workflow.
• Strong understanding of model evaluation, error analysis, precision and recall tradeoffs, and regression testing.
• Ability to reason about ambiguous language and translate linguistic failure modes into engineering mechanisms.
• Strong software engineering fundamentals including testing, Git, Linux, APIs, and maintainable system design.
• Ability to investigate unfamiliar behavior independently and communicate findings clearly.
• A research background is valuable here, but it has to come with evidence that you can build, debug, and own real software systems.

Preferred qualifications

• Relation extraction, named entity recognition, entity linking, or record linkage.
• Retrieval systems, semantic search, reranking, or retrieval augmented generation.
• Fine tuning, supervised instruction tuning, parameter efficient training, or model adaptation.
• Structured generation and constrained decoding.
• Model inference optimization or serving.
• Evaluation against human reviewed truth datasets.
• Financial, legal, scientific, regulatory, or similarly complex documents.
• Knowledge graphs or relationship oriented data.
• Weak supervision, dataset construction, or difficult labeling programs.
• Experience combining deterministic rules with statistical or generative models, or investigating hallucinations, unsupported outputs, attribution errors, or model calibration.