Personal details
Arthur A.
About
Production AI platform engineer, forward-deployed technical leader, and former Microsoft SDE with experience designing and shipping complex agentic systems, RAG platforms, data infrastructure, and regulated financial products.
Built production AI systems from the ground up for legal, industrial, financial, and research-intensive use cases. Recent work includes a LangGraph-based digital analyst with real-time streaming, distributed document ingestion, production tracing, and failure-safe processing; a privacy knowledge platform spanning 127+ jurisdictions; and stateful AI agents reasoning over industrial sensor telemetry.
Combines hands-on Python, Go, FastAPI, LangGraph, Celery, PostgreSQL, Redis, and cloud infrastructure experience with enterprise integration and technical leadership. Particularly effective in environments where AI systems must remain observable, reliable, auditable, and maintainable after the initial demonstration.
While strongest in AI infrastructure, RAG, and agent orchestration rather than marketing automation, brings the platform engineering and integration experience required to build dependable AI workflows for GTM and customer-facing products.
Core Expertise
Agentic AI: LangGraph, LangChain, multi-step agent workflows, tool calling, stateful orchestration, structured outputs, streaming responses
RAG and LLM Systems: Retrieval architecture, hybrid search, document ingestion, citation handling, prompt versioning, LLM evaluation, model routing, LoRA fine-tuning
Languages and Frameworks: Python, Go, C#/.NET, TypeScript, JavaScript, SQL, FastAPI, React
Data and Retrieval: PostgreSQL, pgvector, Pinecone, ChromaDB, FAISS, DynamoDB, Redis, PySpark, Airflow, KQL
Distributed Infrastructure: Celery, RabbitMQ, MassTransit, SQS/SNS, Docker, Kubernetes, Azure, AWS ECS, Lambda and S3
Reliability and Observability: OpenTelemetry, Langfuse, distributed tracing, dead-letter queues, idempotent processing, retry design, regression evaluation, safe rollouts
Domains: Fintech, fraud and risk, pensions and wealth, legal technology, industrial telemetry, enterprise data, GTM-enabling AI
Work Experience
- Designed and built Pangolin AI, the platform behind Attercopās ā24/7 Digital Analyst,ā from the ground up, contributing approximately 50k lines across four core platform layers.
- Designed the agent orchestration layer in LangGraph and bridged graph execution to Server-Sent Events for real-time delivery of intermediate status and generated answers.
- Built a stateful inline-citation parser that preserved citation accuracy while responses were still streaming, rather than correcting references only after generation completed.
- Instrumented individual orchestration nodes with OpenTelemetry and Langfuse, preserving nested traces and prompt-version provenance so failed or low-quality interactions could be reconstructed in production.
- Built a distributed indexing pipeline using Celery and RabbitMQ, with separate I/O, CPU, and image-processing worker fleets, per-stage dead-letter queues, bounded retries, and watermark-based idempotency.
- Introduced cross-document embedding batches that reduced average embedding API latency per chunk by approximately sixfold.
- Built the identity bridge across Sitefinity, VIP, and Pangolin using opaque-token introspection, identity resolution, and tenant-aware Redis session caching.
- Designed a two-layer token and session cache that allowed account top-ups to become visible in under approximately 30 seconds.
- Separated authentication and authorization states, enabling token-exhausted users to retain read-only access and resolving a redirect loop affecting blocked accounts.
- Generated OpenAPI and Celery task references directly from the live route and task graphs, with CI checks that made public-surface changes reviewable before release.
- Owned the production concerns surrounding the AI platform, including streaming behavior, observability, access control, failure recovery, API hygiene, and operational reliability.
LLMs, RAG, Pinecone, FastAPI, DynamoDB, AWS
- Led the design and development of Python and Go services for document understanding and production RAG applications.
- Evaluated Falcon, LLaMA, Code LLaMA, Raven, and other models for latency, cost, stability, and hallucination risk.
- Experimented with memory-efficient LoRA fine-tuning for open-source language and image models.
- Designed vector retrieval across Pinecone, FAISS, and ChromaDB, selecting strategies based on production latency and quality requirements.
- Built session-aware FastAPI services with DynamoDB-backed state and deployed supporting workloads through ECS and Lambda.
- Led the Wall Street Journalās first screen-reader experiment aimed at improving article accessibility.