AI Engineer

ASCENRA TECHNOLOGIES; Varanasi; Hybrid; INR 3,500,000 - 5,200,000; Full time

Job summary

ASCENRA TECHNOLOGIES is seeking an AI Engineer to own the RAG core, ingestion, chunking, embeddings and retrieval over privacy notices and regulatory text. You will simplify legal language for end users and explain what changed when notices are updated.

The role emphasizes grounding, auditability, and guardrails, with a focus on production-grade pipelines and compliance-driven design across multiple languages. Hybrid, full-time role in Greater Noida.

Qualifications

• 4 years of software engineering, or 3 years with an advanced degree — or equivalent practical experience.
• 2 years shipping LLM or RAG features to production and keeping them working — prototypes and notebooks are not the same job
• Strong Python, and the engineering discipline to write services rather than scripts
• Retrieval built properly: defensible chunking strategy, justified embedding model selection, tuned vector search
• An evaluation harness you have built and run — offline evals, regression on prompt changes, honest measurement of grounding and faithfulness
• Prompt engineering as an engineering practice: versioned, tested, and changed on evidence
• Guardrails in production — PII detection and masking, output validation, refusal behaviour, prompt injection as a live threat

Responsibilities

• The RAG core: ingestion, chunking, embeddings, retrieval and reranking over privacy notices, policies and regulatory text
• Notice simplification — turning legal language into something an ordinary person understands without changing what it means
• Version diff: telling somebody in plain language what actually changed when an organisation updated its notice
• Grievance classification and routing, with statutory response clocks running so a misroute has a cost
• Grounding and auditability: source logging, citation, and an evaluation harness that is part of CI
• Guardrails — PII masking, injection defence, and knowing when the right output is a refusal

Skills

Python

LLM features

RAG

Vector search

CI/CD

Prompt engineering

Guardrails

Data privacy

Education

B.E./B.Tech or M.Tech in CS/IT

Tools

pgvector

Job description

AI Engineer

Own the retrieval core and every AI feature in the product: turning dense privacy notices into something a person can understand, showing people what changed when an organisation revises its terms, and classifying and routing grievances within statutory timelines.

Location

Greater Noida

Working Style

• Hybrid
• Full-time, permanent
• Reports to: Head of Engineering

Experience

4+ years

Why this role

Build grounded AI inside a boundary the model is not allowed to cross.

1. Every output grounded in a source we can point at
2. Evaluation harness in CI, not in a spreadsheet
3. Twenty-two languages, with evaluation literature you will have to build
4. Adversarial by default — a consent system is a target

What you’ll own

• The RAG core: ingestion, chunking, embeddings, retrieval and reranking over privacy notices, policies and regulatory text
• Notice simplification — turning legal language into something an ordinary person understands without changing what it means
• Version diff: telling somebody in plain language what actually changed when an organisation updated its notice
• Grievance classification and routing, with statutory response clocks running so a misroute has a cost
• Grounding and auditability: source logging, citation, and an evaluation harness that is part of CI
• Guardrails — PII masking, injection defence, and knowing when the right output is a refusal

The hard part

• The model cannot see the data. Personal data routed through the platform must not be readable by us, which limits what can reach a model provider at all
• Wrong is not a quality metric here — if we simplify a notice and change its meaning, a person consents to something they did not agree to
• Twenty-two languages: notices may be required in English or any language in the Eighth Schedule, across scripts where you will have to build your own measures
• Adversarial by default — assume somebody will try to make your model say the wrong thing on purpose

Minimum qualification

• 4 years of software engineering, or 3 years with an advanced degree — or equivalent practical experience
• 2 years shipping LLM or RAG features to production and keeping them working — prototypes and notebooks are not the same job
• Strong Python, and the engineering discipline to write services rather than scripts
• Retrieval built properly: defensible chunking strategy, justified embedding model selection, tuned vector search. pgvector preferred
• An evaluation harness you have built and run — offline evals, regression on prompt changes, honest measurement of grounding and faithfulness
• Prompt engineering as an engineering practice: versioned, tested, and changed on evidence
• Guardrails in production — PII detection and masking, output validation, refusal behaviour, prompt injection as a live threat

Preferred

• MLOps: model and prompt versioning, drift and cost monitoring, rollback
• Adversarial work on LLM systems — red-teaming, jailbreak resistance, injection defence
• NLP for Indian languages, particularly Indic-script tokenisation and the evaluation problems that follow
• Legal, regulatory or other high-stakes text where being wrong has consequences
• Running open-weight models where data cannot leave the boundary

How we work

• Hybrid working style with anchor at our Grandthum office in Greater Noida, Tech Zone IV
• Defined core collaboration hours — outside that, flex your day around its natural shape
• Architecture decisions get written down and argued in the open. Small, reviewable changes. Depth is valued over volume
• Company-issued devices and approved tooling for anything touching customer or compliance data. Given what we build, we hold ourselves to the standard we sell

The practical details

• Location: Grandthum, Tech Zone IV, Greater Noida West, Gautam Buddha Nagar, Uttar Pradesh
• Employment type: full-time, permanent
• Reports to: Head of Engineering
• Education: B.E./B.Tech or M.Tech in CS/IT, or equivalent demonstrated experience — we mean the second part
• Offers are subject to standard background verification, which we will explain before we ask for anything

How we'll interview you

• A 30-minute conversation with our Head of Engineering
• A code and architecture discussion on something you have built and can talk about honestlyA conversation with the founder about the company, the regulation and where this goes
• Four stages. We aim to complete them inside two weeks and to give you a decision either way

Ideal candidate

You’ll thrive here if this sounds like you

Grounding and evaluation are the parts of this work you actually care about

You can defend compliance properties to people who are not impressed by benchmarks

You architect AI inside hard boundaries and decide what runs where

You treat prompts, evals and guardrails as engineering artefacts under version control

You are comfortable early: eighteen people, pre-revenue, and a product that must be right before a regulator looks at it