LLM · RAG · AGENTS · REMOTE
Ship LLM features that survive production traffic.
RAG pipelines, agentic workflows, voice interfaces, and on-device inference. I build the boring infrastructure that makes the AI feel magical.
·
By Amit Chakraborty — Kolkata, India · Remote worldwide
What you get
- RAG pipelines that actually retrieve — Hybrid search (BM25 + vectors), query rewriting, re-ranking with cross-encoders, and evaluation harnesses so you know when retrieval regresses.
- Agentic workflows — Tool-calling architectures with proper retry, timeout, and human-in-the-loop guardrails. Built on the model layer — not framework lock-in.
- On-device inference — Llama 3, Phi-3, and Gemma running on-device via llama.cpp / MLC for privacy-sensitive use cases.
- Voice interfaces — Whisper STT + ElevenLabs / OpenAI TTS + interruption handling. Sub-second turn-taking latency.
- Eval-driven development — Every LLM feature ships with a regression suite. You should know when GPT-5.2 breaks your prompt before your users do.
Proof
- Kshem (HealthTech RAG) · HIPAA-compliant · sub-2s answer latency — Patient-facing medical assistant with PHI redaction, citation tracking, and an immutable audit trail.
- Aura voice assistant · 740ms turn-taking · barge-in support — Real-time voice agent with ElevenLabs TTS, Whisper STT, and interruption-aware turn detection.
- Internal search agent · 62% → 91% answer accuracy — Hybrid retrieval + cross-encoder re-ranking + query rewriting on a 200K-document corpus.
How it works
- Week 0 — Use-case fit — Free 60-minute scoping call. Half of LLM projects don't need an LLM — I'll tell you that for free.
- Week 1 — Eval harness first — Before writing the production prompt, we build a 50–100 example evaluation set. This is the single best ROI in LLM work.
- Week 2+ — Ship + iterate — Production deployment with logging, tracing (LangSmith / Helicone), and a regression-tested prompt.
Frequently asked questions
Which LLM provider should I use?
It depends on your data sensitivity, latency budget, and cost ceiling. For most B2B apps: Claude 3.5 Sonnet or GPT-5 for quality, Gemini 2.5 Flash for cost. For PHI/PII: Azure OpenAI or AWS Bedrock with a BAA. For privacy-critical: on-device Llama 3.
Do I need a vector database?
Probably yes, but start with Postgres + pgvector before reaching for Pinecone. You can scale to ~1M documents on pgvector with proper HNSW indexing — switch only when you have a real reason.
How do you prevent hallucinations?
You can't eliminate them, you can only design around them. Citation-required answers, structured output validation, and a 'I don't know' escape hatch in the system prompt. And evals — always evals.