Arish K.

I build AI agents and RAG systems that are tested, access-controlled and safe to run with real customers, and I ship them inside full web and mobile products. Every number on this profile comes from a public repo you can run yourself. Most AI projects work in the demo and fail with real users. Nobody measured the wrong answers. The agent repeats a refund after a crash. A tool can read every customer's data. I am hired to prevent that, or to fix it after it happens. WHAT I BUILD Agents you can trust with real actions I write the agent loop and the harness around it myself, in TypeScript or Python. Every tool call is validated, traced and capped by a budget. Risky actions wait for a person's approval. If the server dies mid-task, the run resumes and nothing happens twice. Proof: a customer-service agent scored on the public tau-bench airline and retail tasks. It passes 45% of tasks on one try and 20% on four tries in a row, at $0.08 per task. A plain scripted workflow scores 35% on the same tasks. In 100 forced crashes, 0 actions ran twice and 0 ran without approval. Evals, tracing and cost control for an agent you already have A test set built from your real failures, a release gate that blocks a deploy when quality drops, a trace for every run, and spending limits that actually stop a run. Proof: prompt caching cut my cost per task from $0.12 to $0.07 with the same pass rate. I publish the failure categories with counts, including the fixes that did not work. Secure MCP servers and permission layers OAuth 2.1 sign-in, a per-user and per-customer check on every tool call, pinned tool definitions, sandboxed code execution, personal data masked in logs, and a full audit trail. Proof: 0 of 500 attempts to reach another customer's data succeeded. On the AgentDojo prompt-injection tests, attack success fell from 40% to 12% while 90% of normal tasks still passed. RAG that answers correctly, cites its source and says when it does not know Keyword plus vector search, reranking, parsers that keep tables intact, citations, and search that respects who may see each document. Proof: recall@10 rose from 0.72 to 0.86 on a 200-question custom test set, measured one change at a time. One popular change, adding a reranker, made results worse on this data, and the write-up says so. Voice agents The same agent on a phone line, built two ways and compared: 1.2 seconds to first audio and $0.08 per call for the pipeline design, 0.7 seconds and $0.12 per call for the realtime design. SHIPPED TO REAL USERS • AI tutor and document-to-quiz features for an exam-prep app: retrieval, streaming answers, OCR, queue workers and rate limits, used by real students. • An MCP client with OAuth 2.1 and PKCE, tokens in the system keystore, a tool allow-list and an audit log. • 8+ years of full-stack and mobile delivery in React, Next.js, Node, NestJS, Go, Python and Flutter, including about 100 projects shipped as CTO of my own software company. HOW I WORK You get a baseline number in the first week. After that you get a short report every week: pass rate, cost per task, and what still fails. I will tell you when a plain workflow beats an agent, because it often does. I work 40 hours a week, with 4 hours of overlap with US Eastern time. STACK TypeScript, Python, Go. OpenAI and Anthropic Claude APIs. MCP TypeScript SDK. Postgres with pgvector. Langfuse and OpenTelemetry. Docker. AWS. LangGraph when your codebase already uses it. Send me the job post, or the thing that is breaking, and I will tell you how I would measure it first.