Principal Engineer

Truemeds; Mumbai; On-site; INR 4,000,000 - 7,000,000; Full time

Job summary

Truemeds seeks a senior AI technologist to conceive, build, and ship AI products, ML models, RAG systems, agents, and decisioning engines that impact the business. You will work with the founder’s office on all things AI, avoiding platform-architecture or backend-only roles.

You will own end-to-end AI initiatives, from problem framing to production, with direct access to budget, and a mandate to move fast, test ideas, and measure outcomes across the P&L and innovation portfolio.

Qualifications

• Production-ready GenAI systems with measurable impact.
• Prompt engineering, structured outputs, and tool calling.
• RAG end-to-end deployment with retrieval quality metrics.
• Able to measure unit economics: cost per call and latency.
• Produce end-to-end ML systems from features to online evaluation.

Responsibilities

• Take ambiguous business problems and ship a working AI system (POC-first, pilot, scale).
• Ship GenAI at production depth: LLM apps, RAG, agentic systems, evals, guardrails, latency, cost per call.
• Take ML models to production across features, training, offline/online eval, deployment and drift.
• Own highest-upside P&L bets and an active innovation portfolio; kill fast, double down on what works.
• Set AI technical bar: architectures, tooling, eval standards, cost governance, safe-AI in healthcare data.

Skills

GenAI leverage

Prompt engineering

RAG & tool calls

Cost per task mgmt

End-to-end ML prod

LLM deployment

Agent frameworks

Tools

PyTorch

HuggingFace

OpenAI API

Claude / Gemini

Pinecone

Qdrant

Weaviate

Milvus

Chroma

FAISS

pgvector

LangGraph

AutoGen

CrewAI

Docker

Kubernetes

MLflow

Weights & Biases

SageMaker

Kubeflow

TensorRT-LLM

Job description

Truemeds is making healthcare affordable for every Indian. FY26 is when AI stops being a tool and becomes an operating capability across Marketing, Engineering, Data, Product and CX. We want the technologist who conceives, builds and ships the AI products, ML models, RAG systems, agents and decisioning engines that move the company, working as a parallel to the founder's office on all things AI. To be clear about what this is not: it is not a platform-architecture or backend-leadership seat. If your first instinct on hearing “scale” is service meshes and Kafka topics, this isn't your role. If it is retrieval quality, eval harnesses and cost per call, keep reading. Any stack, any model, a real budget, one bar: ship AI that visibly changes the business.

Why it's different

• 1.Horizontal mandate. You lead AI-led technology growth for the whole company, not one squad.
• 2.Not BAU. You build the 0-to-1 bets, models and platforms that other teams then consume.
• 3.Roughly half your work is direct P&L (margin, retention, conversion, cost-to-serve), and roughly half is pure innovation (moonshots and POCs with no line item yet).
• 4.Direct founder access, the shortest path from idea to go/no-go, and total stack freedom, judged on outcomes rather than orthodoxy.

What you'll do

• 1.Take an ambiguous business problem, across growth, supply chain, customer experience, pricing, operations and risk, and ship a working AI system: POC-first, scoped pilot, clear go/no-go, then scale.
• 2. Ship GenAI at production depth: LLM apps, RAG, agentic and tool-use systems, evals, guardrails, latency and token economics.
• 3.Take ML models to production across the full loop, from features and training to offline and online evaluation, deployment and drift.
• 4.Own the highest-upside P&L bets and a live innovation portfolio. Kill fast, double down on what works.
• 5.Set the AI technical bar (reference architectures, tooling, eval standards, cost governance, safe-AI) and build compliance-aware by default, given healthcare data, Rx and PII.

Non-negotiable skills, proven, not aspiring

We hire on evidence. Point to production systems you personally built, and the numbers you moved.

• 1. GenAI, as leverage. You turn LLM APIs (OpenAI, Anthropic/Claude, Gemini, open models via HuggingFace) into force, orchestrating agents to extract the output of a whole engineering pod, solo. Fluent in prompt and context engineering, structured output, and function/tool calling. You have shipped RAG end to end and you tune retrieval to real numbers, recall@k, MRR and nDCG, with chunking, reranking and hybrid dense plus BM25, on real vector databases (pgvector, Pinecone, Qdrant, Weaviate, Milvus, Chroma, FAISS). You hold generation to faithfulness, groundedness and a low hallucination rate. You build agents (LangGraph, AutoGen, CrewAI or equivalent) and measure them by task success rate and tool-call accuracy. And you know your unit economics cold: cost per resolved task, tokens per request, p95 latency. If you can't measure it, you haven't shipped it.
• 2. Applied ML and Data Science in production. Strong Python DS stack (PyTorch and/or HuggingFace Transformers, scikit-learn, Pandas, NumPy, SQL). Real models live in production: ranking, recommendation, classification, forecasting, propensity and uplift, evaluated on the right metrics (AUC/PR, calibration, offline and online), with A/B testing and drift monitoring. You frame fuzzy problems as ML problems, and you know when not to use ML.
• 3. Productionizing AI and MLOps. Model serving (vLLM, TGI, TensorRT-LLM) with latency optimization, and the techniques to pay for scale: quantization, distillation, speculative decoding and caching. Experiment tracking (MLflow, Weights & Biases), CI/CD for models, Docker and Kubernetes, pipelines (SageMaker, Vertex or Kubeflow).
• 4. Engineering and architecture mastery, you ship whole products alone. Deep in at least one language with the CS fundamentals under it: data structures, algorithms, complexity, concurrency and memory (depth AI can't fake). You architect distributed, scalable, secure systems. Full-stack range: you own the backend, a frontend, the data layer (relational plus NoSQL or vector) and cloud deploy (AWS, GCP or Azure, with scale, security and cost control).
• 5. Proof, and a conversation slideware won't survive. Production systems people actually use, and substantial open-source, published models or public technical work (a competitive-coding record or hackathon wins are a plus, not a gate). In the interview you'll walk us through one AI system you built end to end: the eval harness, the failure modes you found, the cost per call, and what you cut to ship it.

Nice to have

• fine-tuning and adaptation (LoRA, QLoRA, PEFT)
• e-commerce or healthcare experience with regulatory awareness (NDHM, HIPAA-inspired, PCI-DSS)
• having stood up an ML or LLM platform from scratch

Founder mentality, a one-person startup inside Truemeds

This is a grind-till-you-make-it seat. No playbook, no team, no hand-holding. You scope it, build it, ship it, measure it and own the outcome. Ambiguity and total freedom energize you, and you'd rather ship a prototype today than a perfect plan next quarter. You build with agentic tools (Claude Code, Cursor, Windsurf) reflexively, and it shows in your velocity.

The liberty and the budget Any stack, any model, your call. A real innovation budget to spin up compute, tools, APIs, vector stores and POCs without fighting for every rupee. Direct founder and leadership access. And a team-building mandate when you're ready: start as a pure IC and grow into founding and leading an AI pod.