AI & LLM Integration (RAG, Agents, Cost-Aware Routing)Pratyush Sudhakar
I build AI features that hold up in production, not demos that break on the second prompt. Retrieval that returns the right chunk, agents that remember across sessions, and routing that keeps your inference bill from surprising you.
Recent work: a RAG pipeline over course syllabi answering natural-language queries at roughly 50ms average latency. A self-hosted orchestration stack routing across 12+ LLM providers with tiered fallbacks, cutting routine inference cost by 30 to 150x. An infinite canvas that classifies and links freeform notes in real time using GPT-4 and embeddings.
What I build
Chat over your documents: ingestion, chunking, embeddings, reranking, and evaluation so answers stay grounded
LLM features inside an existing product, with streaming, retries, and graceful degradation when a provider fails
Agent systems with persistent vector memory and typed tool calls
Multi-provider routing with ordered fallback chains, budget guardrails, and separate handling for context-window failures
Observability you can actually read: token spend, latency, and failure traces
How I approach it
Measure before optimizing. On one pipeline that meant finding the bottleneck was input size, not the model, and cutting latency from 3 minutes to under 50 seconds
Cheapest acceptable model as the default, expensive models as fallbacks, never the other way around
Evals before launch, so you know when a prompt change makes things worse
Pricing
Hourly from $75/hr
Fixed-scope builds from $3,000, quoted after a scoping call
Free 20-minute call to size the problem before anyone commits
AI & LLM Integration (RAG, Agents, Cost-Aware Routing)Pratyush Sudhakar
I build AI features that hold up in production, not demos that break on the second prompt. Retrieval that returns the right chunk, agents that remember across sessions, and routing that keeps your inference bill from surprising you.
Recent work: a RAG pipeline over course syllabi answering natural-language queries at roughly 50ms average latency. A self-hosted orchestration stack routing across 12+ LLM providers with tiered fallbacks, cutting routine inference cost by 30 to 150x. An infinite canvas that classifies and links freeform notes in real time using GPT-4 and embeddings.
What I build
Chat over your documents: ingestion, chunking, embeddings, reranking, and evaluation so answers stay grounded
LLM features inside an existing product, with streaming, retries, and graceful degradation when a provider fails
Agent systems with persistent vector memory and typed tool calls
Multi-provider routing with ordered fallback chains, budget guardrails, and separate handling for context-window failures
Observability you can actually read: token spend, latency, and failure traces
How I approach it
Measure before optimizing. On one pipeline that meant finding the bottleneck was input size, not the model, and cutting latency from 3 minutes to under 50 seconds
Cheapest acceptable model as the default, expensive models as fallbacks, never the other way around
Evals before launch, so you know when a prompt change makes things worse
Pricing
Hourly from $75/hr
Fixed-scope builds from $3,000, quoted after a scoping call
Free 20-minute call to size the problem before anyone commits
$75 /hr