Amir – OpenAI, Python, RAG

An agentic AI assistant embedded in the policyholder app of a major insurance group. Car policyholders need fast, reliable answers about whether a claim is covered, and an incorrect "covered" answer carries direct financial and regulatory risk for the insurer. The assistant answers coverage questions from each customer's structured contract data, routes every message to a specialized agent, and hands over uncertain cases to human advisers. It runs on a multi-agent orchestrator with a multi-provider LLM gateway and shared guardrails and security checks across all agents.

• Led end-to-end the build of the claims-coverage agent for car policyholders, raising answer accuracy from 85% to 96% on a 203-question control set with zero claims wrongly marked as covered.
• Implemented a code-level guard that intercepted any "covered" answer on a guarantee missing from the customer's contract, prioritizing reliability over coverage and routing uncertain cases to human advisers.
• Defined the evaluation method with business teams and built the control set (203 questions, later extended to 300+), adopted as the project's reference for every release.
• Designed the multi-agent orchestrator from the prototype: an LLM router, specialized agents with per-agent tool scoping, and a multi-provider LLM gateway.
• Deployed the system to pre-production with shared guardrails and security checks across agents.
• Instrumented the system with OpenTelemetry to track accuracy and p90/p99 latency, and maintained 500+ automated tests.
• Delivered a strategic AI-in-insurance study to the client's leadership.

An applied-AI studio that designs and ships production generative-AI systems for startups, SMEs, and large enterprises that want to automate manual, knowledge-heavy work. Its projects include operational agents that replace manual report review, retrieval-augmented assistants that leverage company knowledge, MCP servers and finance agents connected to trading and accounting tools, and multimodal voice and chat bots. The studio also runs AI audits and in-company AI training programs for corporate teams.

• Put into production a LangGraph agent that read night-patrol reports and called the on-call manager within a minute, saving an hour of manual review every morning.
• Shipped an MCP server used by real traders and a finance agent built on the Pennylane accounting platform.
• Built a multi-tool conversational agent with Qdrant vector retrieval, persisted Postgres conversation memory and external tool calls.
• Built RAG pipelines with Supabase vector search and Cohere reranking, plus a scalable indexing pipeline covering extraction, semantic chunking, embedding generation, and batch upserts.
• Delivered an end-to-end knowledge-processing pipeline with 40+ steps, conditional routing, error handling, and recovery.
• Built multimodal voice and chat bots with ElevenLabs and orchestrated multi-agent systems with LangGraph and MCP servers.
• Ran 8 AI audits of SMEs, scoped and estimated AI projects, and delivered in-company AI training for world-renowned brands

A secure retrieval-augmented knowledge assistant for the R&D division of a global energy company. More than 1,000 research documents were siloed across 10+ innovation labs, so teams and leadership could not see or reuse each other's work. The assistant indexes the full R&D corpus, answers questions with retrieval restricted by per-document access rights, and generates on-demand summaries and one-page briefs from technical reports. It runs in a secure Azure environment to meet the company's security and document governance requirements.

• Built and deployed a production RAG system over 1,000+ R&D documents across 10+ innovation labs, reducing document search time by 68% on the measured scope and giving leadership visibility into every lab's work for the first time.
• Selected Azure OpenAI and Azure AI Search to meet security, scalability, and document-governance constraints.
• Implemented per-document access rights so each user only retrieved the documents they were allowed to see.
• Built the document ingestion, vector indexing, contextual retrieval, and on-demand summary pipeline.
• Automated the synthesis of 20+ technical reports into one-page briefs, speeding up cross-team decision-making.
• Standardized AI workflows across labs and trained business users on the new tool.

A market-making and liquidity-monitoring service for Web3 and crypto trading firms. Quant teams need clean, low-latency, tick-level market data from multiple exchanges to price assets and monitor liquidity. The system ingests real-time exchange streams, normalizes order-book events into business models, and delivers tick-level feeds to quant teams, with monitoring for feed outages and scheduled exchange maintenance.

• Built a real-time market-data stream manager in Rust, mapping order-book events to business models with sub-10 ms latency and delivering tick-level feeds to quant teams in production.
• Created Python monitoring scripts that detected feed outages and alerted teams about scheduled exchange maintenance.
• Wrote the architecture documentation for the feed handler to support handover and future evolution.