AI / ML development for Indian teams.
In production, not in a Jupyter notebook.
Most AI projects in India never reach production — they get stuck in POC purgatory because the team builds the demo and never the boring 70%: data pipelines, evals, monitoring, cost controls, prompt versioning. We start at "this needs to run for real users at real cost" and work backwards.
Serving: Pan-India · Global remote
Concrete deliverables for India (remote) clients.
- LLM application engineering — RAG, agentic workflows, function calling, structured outputs
- Fine-tuning when it actually wins — LoRA / QLoRA on Llama 3, Mistral, etc.
- Vector database + retrieval — Pinecone, Weaviate, pgvector, Qdrant
- Traditional ML — forecasting, classification, recommendation, anomaly detection
- MLOps — experiment tracking (MLflow / W&B), model registry, CI/CD for models
- Evals + observability — Langfuse / Helicone / custom; quality gates in CI
- Cost controls — prompt caching, request routing (cheap → expensive models), token budgets
- GPU infrastructure — managed (SageMaker, Vertex) or self-hosted (EKS + Karpenter + GPU nodes)
Local context, not boilerplate.
The Indian AI/ML market is split between "agencies that build POCs and disappear" and "global vendors at $300/hour". We sit in the middle: senior engineers who actually ship, INR pricing, and we stay engaged for 3+ months to operate what we build.
For Indian SaaS teams, the highest-value AI projects right now are: customer support augmentation (RAG over docs/tickets), sales rep enablement (search + generation over CRM), document processing (invoice/KYC/contract OCR + extraction), and code/content moderation. We've shipped each pattern.
For traditional ML — forecasting, churn, recommendation — we still default to gradient-boosted trees (XGBoost / LightGBM / CatBoost) over deep learning unless data volume and pattern complexity justify it. Most Indian SMEs do not have the data volume to make deep learning the right answer; we tell you that up-front.
Questions India (remote) clients ask first.
- Should we use OpenAI / Anthropic / open-source models?
- Honest answer first. For most use cases, Anthropic Claude or OpenAI GPT-4-class models hit production fastest and are cheapest at low-to-moderate volume. Self-hosted (Llama 3 / Mistral) wins above ~5M tokens/day or when data residency forces it. We design for portability — same prompts, swappable backend.
- Do you do RAG?
- Yes — we've shipped RAG systems for customer support, internal search, sales enablement and legal document retrieval. We're opinionated about hybrid retrieval (BM25 + vector + reranking) and aggressive eval before production.
- How do you handle data residency for Indian clients?
- For India-only data, we keep everything (vector DB, model inference, logs) in ap-south-1. For LLM calls — Anthropic and OpenAI offer EU/US-only routing; for strict data residency we self-host on Indian GPU infra. We design this into the architecture, not bolt it on.
- What about hallucinations in production?
- Mitigated via three layers: retrieval grounding (RAG with source citation), structured outputs with schema validation, and a dedicated eval suite that runs in CI (you cannot deploy a prompt change that regresses the eval set). For high-stakes outputs we add a human-in-the-loop checkpoint.
- Can you fine-tune models?
- Yes — LoRA / QLoRA on Llama 3 / Mistral / Phi. We recommend fine-tuning only when prompt engineering + RAG don't close the gap, because fine-tuning is harder to maintain. Roughly 20% of AI engagements end up needing fine-tuning.
Same team, other cities and tracks.
Talk to a senior engineer.
30-min architecture review · written assessment within 48h · no commitment.