Artificial Intelligence Engineer (New Delhi)

Artificial Intelligence Engineer (New Delhi)

19 Nov
|
Yogesher India
|
New Delhi

19 Nov

Yogesher India

New Delhi

Why this role exists

- Our existing AI work needs a single accountable engineer for design, monitoring, cost and quality.
- Every AI feature touches protected health information (PHI) and some influence clinical or financial decisions, so engineering discipline and compliance are not optional.
- We want someone who knows Claude deeply and also understands the wider AI landscape well enough to choose the right tool, including when that tool is not an LLM.

What you will do

Build

- Design and ship Claude-powered features: ambient or dictated visit-note drafting, chart summarization, patient message triage and reply drafts, prior-authorization and appeal letters, referral summaries.
- Build revenue cycle automation: charge and code suggestions (ICD-10, CPT, HCPCS, modifiers), claim scrubbing, denial analysis and appeal drafting, eligibility and remittance (835) reconciliation, A/R follow-up prioritization.
- Build agentic workflows with tool use and Model Context Protocol (MCP) servers that read from and write to the EHR and billing database through narrow, audited, least-privilege interfaces.
- Build retrieval (RAG) over clinical guidelines, payer policies, fee schedules and internal SOPs, with citations back to the source.
- Integrate with the EHR, practice management, clearinghouse (Waystar) and patient portal through FHIR R4, HL7 v2, X12 and internal APIs.

Run and maintain

- Own the existing AI codebase, prompts, tools and pipelines: document them, refactor them, and put them under version control with tests.
- Build and run an evaluation suite (golden sets, regression tests, LLM-as-judge with clinician spot checks) that gates every prompt, model or tool change.
- Monitor quality, latency, error rates, refusals, hallucination and drift in production; set alerts and on-call runbooks.
- Manage model selection, upgrades and deprecations across Claude tiers; keep prompts portable and rerun evals before every migration.
- Track and control spend: prompt caching, batch processing, model routing, token budgets, per-feature cost dashboards.

Govern and comply

- Keep PHI safe: minimum-necessary data in prompts, de-identification where possible, no PHI in logs or traces, retention controls, and a signed Business Associate Agreement (BAA) with every vendor that touches PHI.
- Support ONC certification: maintain documentation, source attributes and risk management for any AI that qualifies as a decision support intervention, and contribute to real-world testing and surveillance.
- Run AI risk assessments, bias and safety reviews, and incident response for AI errors, with clinical sign-off before launch.
- Keep an inventory of every AI use case, model, prompt version, data flow and owner.

Collaborate

- Work with physicians, nurses, coders, billers and front-desk staff to find the highest-value use cases and to test outputs.
- Work with compliance, privacy and security officers on policies, audits and vendor reviews.
- Train staff on safe AI use, explain limits plainly, and collect feedback.

Required technical skills Claude and LLM engineering (must have)

- Production experience with the Claude API: Messages API, system prompts, multi-turn state, streaming, structured and JSON outputs, tool use and parallel tool calls, extended thinking, vision and PDF input, citations, Files API, Batch API, and prompt caching.
- Model selection across the Claude family (fast/cheap versus most capable), with a measured cost, latency and quality tradeoff for each use case.
- Building agents with the Claude Agent SDK or an equivalent loop, including permissions, sandboxing, human-in-the-loop approval and failure recovery.
- Writing and securing MCP servers and clients; knowledge of Claude Code and Skills for engineering productivity.
- Access paths and their compliance differences: Anthropic API, AWS Bedrock, Google Vertex AI; zero-data-retention and BAA terms.
- Advanced prompt engineering: role and context design, few-shot examples, XML-style structure, long-context handling, prompt versioning and A/B testing.
- Defending against prompt injection and data exfiltration, especially from documents, faxes, patient messages and web content that flow into prompts.

Retrieval and data (must have)

- RAG design: chunking, embeddings, hybrid search, reranking, citation grounding, freshness and access control per user or role.
- Vector search (pgvector or equivalent) and full-text search; document parsing and OCR for scanned records and faxes.
- Strong SQL and relational modeling (PostgreSQL required; SQL Server a plus). Query tuning, indexing, migrations.
- ETL and data quality pipelines; handling of clinical and billing data with audit trails.

Evaluation and quality (must have)

- Designing eval sets from real, de-identified cases; metrics for accuracy, groundedness, completeness, harmful-output rate and code or billing-field correctness.
- Automated regression testing in CI, LLM-as-judge with calibration against human reviewers, and online monitoring with sampled clinician review.

Software engineering (must have)





- Strong Python and one of C#/.NET or TypeScript; our core applications are in the .NET and PostgreSQL ecosystem, so comfort reading and integrating with that code is expected.
- REST, WCF/SOAP and message-queue integration; async processing, retries, idempotency, rate-limit handling.
- Git, code review, automated tests, CI/CD, containers, infrastructure as code.
- Cloud (AWS, Azure or GCP) in a HIPAA-eligible configuration: VPC isolation, KMS encryption, secrets management, IAM least privilege.

Observability and operations (must have)

- Structured tracing of prompts, tool calls and outputs with PHI redaction; dashboards for cost, latency, error and quality; alerting and runbooks.

Security (must have)

- Encryption in transit and at rest, role-based access, audit logging, secure SDLC, dependency and vulnerability management, threat modeling for LLM applications (OWASP LLM Top 10).

Required healthcare and compliance knowledge The candidate must be able to apply these, not just name them. Priority: Must = required at hire, Learn = acceptable to learn in the first 90 days.

Area

What the candidate should know

Priority

HIPAA Privacy, Security and Breach Rules

PHI and minimum necessary, BAAs, safeguards, audit controls, de-identification (Safe Harbor and Expert Determination), breach notification, AI-specific risks such as PHI leaking into prompts and logs

Must

ONC Health IT Certification

Certification criteria our product holds, Cures Act information blocking rules, real-world testing, and the HTI-1 final rule requirements for decision support interventions (170.315(b)(11)), including source attributes, risk management and transparency for predictive models

Must

FHIR and interoperability

FHIR R4, US Core, USCDI, SMART on FHIR, OAuth 2.0, the certified standardized API (170.315(g)(10)), C-CDA, bulk data; HL7 v2 messages (ADT, ORM, ORU, SIU)

Must

Billing and revenue cycle

X12 837, 835, 270/271, 276/277, 278; claim lifecycle, denials and CARC/RARC codes, clearinghouse workflows, payer rules, patient statements and payment posting

Must

Medical coding

ICD-10-CM, CPT, HCPCS, modifiers, E/M leveling, medical necessity, NCCI edits, risk adjustment basics

Must

Clinical workflow

Visit documentation (SOAP), orders and results, e-prescribing, medication reconciliation, problem list, referrals, prior authorization

Must

Clinical terminologies

SNOMED CT, LOINC, RxNorm, CVX, UCUM; mapping free text to codes

Learn

FDA and clinical decision support

When software is a regulated device and the CDS exemption criteria; FDA guidance on AI-enabled software

Learn

Nondiscrimination and bias

Section 1557 patient care decision support tools, bias testing across demographic groups

Must

State and payer AI rules

State laws on AI disclosure in patient communication, telehealth and privacy; payer rules on AI in utilization review and prior authorization

Learn

Security frameworks

NIST CSF, NIST AI RMF, HITRUST or SOC 2 awareness, 42 CFR Part 2 for substance use records

Learn

Patient safety

Escalation design, clinician-in-the-loop, error reporting, and the difference between administrative and clinical AI risk

Must

Broader AI knowledge expected

We use Claude, but the engineer must see the whole field so we pick the right tool and are never locked in by habit.

- Other LLM providers and open models: working knowledge of OpenAI, Gemini and open-weight models (Llama, Mistral, Qwen), and when self-hosting is justified for PHI, cost or latency reasons. Ability to run a fair head-to-head bake-off.
- Classical machine learning: supervised learning, gradient boosting, calibration, and validation for risk scores, no-show prediction, denial prediction and propensity models; scikit-learn, XGBoost or similar.
- NLP and information extraction: clinical named entity recognition, de-identification models, classification and entity resolution; when a small fine-tuned model beats a large LLM.
- Speech and ambient scribing: speech-to-text (Whisper-class and commercial medical ASR), diarization, medical vocabulary handling, and consent and recording-law considerations.
- Document AI: OCR, layout analysis, fax and scanned-form ingestion, insurance card and EOB extraction.
- Multimodal models: image and PDF understanding; clear limits on any diagnostic imaging use, which is out of scope without regulatory review.
- Fine-tuning and adaptation: when fine-tuning, distillation or retrieval is the right lever, and how to validate each.
- Responsible AI: hallucination and automation-bias mitigation, explainability suited to clinicians, fairness testing, model cards, drift monitoring, and incident reporting.




- Standards and frameworks: NIST AI Risk Management Framework, ISO/IEC 42001 awareness, CHAI and Joint Commission guidance on responsible AI in health care.
- Awareness of the market: commercial ambient scribe, coding and revenue cycle AI vendors, so we can judge build versus buy.

Experience, education and personal qualities

Experience

- 4+ years of professional software engineering, including 2+ years building and operating LLM or ML systems in production. [Adjust for seniority]
- At least one shipped LLM application with real users, an eval suite and monitoring.
- 2+ years in health care, health IT or another regulated domain (finance, insurance), with direct handling of PHI or equivalent sensitive data.
- Hands-on integration with an EHR, practice management or billing system.

Education

- Bachelor's degree in computer science, engineering, informatics or a related field, or equivalent experience. A master's degree or health informatics certificate is a plus.

Nice to have

- Experience with an ONC-certified product or the certification process, or with a Waystar or similar clearinghouse integration.
- Clinical or coding credentials (CPC, CCS, RHIA, RN) or prior work alongside clinicians.
- HIPAA, security or AI governance certifications (CHPS, CISSP, Security+, AIGP).
- Experience migrating legacy .NET or SQL Server systems to PostgreSQL.
- Open-source or published work on LLM evaluation, safety or health AI.

Personal qualities

- Transparent communicator who can explain model limits to a physician and a biller in their own language.
- Careful and skeptical: treats every AI output as a claim to be verified, and says no to unsafe shortcuts.
- Pragmatic: prefers the simplest solution that works, and measures before and after.
- Comfortable with ownership, documentation and on-call, in a small team without a large platform group.
- Ethical and discreet with patient information.

First 90 days and success measures

Period

Outcomes

Days 1 to 30

Complete HIPAA and security training. Inventory every existing AI use case, prompt, tool, data flow and vendor agreement. Read the codebase and shadow clinicians and billers. Deliver a gap report covering evals, monitoring, PHI handling and cost.

Days 31 to 60

Put existing prompts and tools under version control with tests. Stand up a baseline eval suite and production dashboard for the top two AI workflows. Fix the highest-risk PHI and security gaps.

Days 61 to 90

Ship one measurable improvement or new feature through the full process (eval, clinical sign-off, staged rollout). Publish the AI governance checklist and incident runbook. Present a 12-month AI roadmap with build-versus-buy and cost estimates.

Ongoing success measures

- Every production AI feature has an owner, an eval suite, monitoring and a documented risk review.
- Zero PHI incidents attributable to AI systems.
- Measured gains, such as minutes saved per visit note, clean-claim rate, denial rate, days in A/R and staff satisfaction, tracked against a pre-AI baseline.
- Cost per task stays within agreed budgets, and model upgrades ship without quality regressions.
- Clinician and staff trust, measured by periodic survey and sampled review.

Hiring process and screening

Stage

What we assess

1. Application screen

Resume against must-have skills; a short written answer on a past LLM system they shipped and how they measured it.
1. Recruiter call (30 min)

Motivation, work authorization, compensation, availability.
1. Technical interview (60 min)

Claude API depth, tool use and MCP, RAG design, prompt injection, cost control, eval design.
1. Practical exercise (take-home 3 to 4 hours, paid, or live 90 min)

Given de-identified sample notes and a claims file, design and partly build a small Claude workflow with an eval set and a PHI-safety plan.
1. Healthcare and compliance interview (45 min)

HIPAA scenarios, ONC decision support requirements, FHIR and X12 fluency, incident handling.
1. Clinical and billing stakeholder panel (45 min)

Communication with a physician and a billing lead; how they would earn trust.
1. References and background check

Standard checks, plus an OIG exclusion screen. Sample screening questions

1. A clinician says the note summary missed a medication. How do you find the cause, fix it, and prove it stays fixed?
2. How would you let an agent update a billing record without risking wrong or unauthorized writes?
3. Which data may go into a prompt, which may not, and how do you enforce that in code?
4. A new Claude model version is released. How do you decide whether and when to migrate?
5. When would you use a rules engine or a small classifier instead of an LLM?
6. What must we document if an AI tool qualifies as a decision support intervention under HTI-1?

Scoring rubric: rate each dimension from 1 to 4: Claude and LLM engineering, evaluation rigor, healthcare and compliance, software engineering, communication. Require 3 or higher in compliance and evaluation for any offer. Compensation and how to apply:- [email protected]

📌 Artificial Intelligence Engineer (New Delhi)
🏢 Yogesher India
📍 New Delhi

Reply to this offer

Impress this employer describing Your skills and abilities, fill out the form below and leave Your personal touch in the presentation letter.

Subscribe to this job alert:

Get the latest job offers by email for: artificial intelligence engineer (new delhi) / new delhi