Amir Azadfar

• Architected and deployed a production-grade, microservice-based AI platform composed of a provider-agnostic LLM client, DAG-based orchestration runtime, and FastAPI services, enabling scalable multi-agent workflows via RESTful APIs and service-oriented architecture (41 domain-specific operators).
• Designed a contract-driven operator framework using JSON manifests, Jinja2 prompt templates, and JSON Schema validation, enabling hybrid deterministic/LLM execution, policy enforcement, human-in-the-loop gates, and resilient operator-level retries, circuit breakers, and rollout controls.case study
• Built and deployed a distributed conversational AI system with 20+ integrated tools, exposing real-time inference via streaming APIs (SSE) and supporting multi-provider LLM execution across OpenAI, Anthropic, and Gemini.case study
• Engineered a production-scale backend system for automated outreach, integrating recommendation pipelines, multi-agent workflows, and asynchronous task processing with external APIs (Gmail OAuth) and persistent state management.
• Built an AI-powered faculty crawling and digestion platform (CanSpider - Professors) that replaced manual collection workflows, expanding coverage from 18,000 professors across 350 departments and 31 institutions in 4 months to 64,127 professors across 2,325 departments and 99 institutions in 40 days through LLM-generated crawl plans, Scrapy/Playwright execution, Kafka pipelines, and review tooling.
• Developed internal operational interfaces using Next.js (React, Tailwind CSS) to manage AI workflows, orchestration pipelines, and system state, enabling real-time control and monitoring of production processes.
• Designed and implemented event-driven pipelines using Kafka for asynchronous data ingestion, processing, and coordination between AI services, improving system scalability and fault tolerance.
• Developed a second AI extraction platform for academic program intelligence (CanSpider - Programs), combining autonomous URL discovery, static/dynamic rendering detection, LLM-based structured extraction, OCR fallback, confidence-weighted aggregation, and config-driven enrichment pipelines for tuition, deadlines, admissions, and language requirements.
• Engineered an AI-powered academic program recommendation system combining semantic vector search with dynamic rule-based multi-filter matching via Qdrant. Designed FastAPI endpoints for real-time student-to-program personalization.
• Built and deployed AI data products including an immigration news intelligence system (CanNews) integrating 23 sources, OpenAI-powered summarization, Elasticsearch semantic search, REST APIs, and Telegram delivery pipelines.case study
• Owned end-to-end system deployment and operations, including Dockerized microservices, Nginx routing, CI/CD pipelines, cloud provisioning (AWS/GCP), monitoring and alerting, caching layers (Redis), and production reliability.

• Built large-scale financial data pipelines aggregating technical, fundamental, and news data for 400+ Tehran Stock Exchange equities; extended infrastructure to crypto markets (Binance, KuCoin) with real-time WebSocket ingestion (<500 ms latency).
• Designed algorithmic trading strategies (mean reversion, momentum, order-book microstructure) and implemented derivatives pricing models including Black-Scholes and binomial trees.
• Developed machine learning models for market segmentation using consensus clustering (K-Means, hierarchical) on historical financial time series.
• Built a Persian (Farsi) NLP sentiment analysis pipeline by labeling 15K+ financial text samples and training TF-IDF + Naive Bayes classifiers used in production analytics.
• Architected a Telegram-based financial analytics platform serving 10K+ users with REST APIs, asynchronous processing, real-time trading signals, and sentiment-driven alerts.
• Developed automated trade execution modules, realistic backtesting environments, and a high-frequency crypto triangular arbitrage system scanning 1000+ trading pairs.
• Deployed and maintained production systems using FastAPI, MongoDB, Redis, Docker, and AWS EC2.

Built a crawling platform where the model writes the crawl plan and the system refuses to trust it: plans are statically linted, dry-run against the live page with per-field coverage reporting, and auto-repaired before execution. One generic Scrapy + Playwright spider replaces N hand-written ones. 64,572 researcher records across 143 institutions in 7 countries from 209,753 crawled pages, at roughly CAD 0.03 per record.

Async runtime kernel for LLM and agent applications: 45,598 lines across 95 modules, 492 tests, and 17 documented stable API namespaces under a published stability policy. Provider adapters behind one Protocol, an execution engine with retries, failover, circuit breakers and idempotency, structured-output validate-and-repair with attempt traces, four cache backends, a USD budget ledger, and deterministic replay. Built deliberately instead of adopting LangChain.

Built a graph neural model to predict drug interaction severity using SMILES-based molecular graphs and RDKit descriptors. Modeled 16,000+ pharmacological structures with GAT, GIN, and MPNN variants to benchmark predictive performance.

Built an end-to-end BTC/USDT futures forecasting and trading pipeline combining Amazon Chronos-2, LightGBM quantile models, regime-aware strategy logic, realistic cost modeling, and walk-forward backtesting/paper-trading for deployment evaluation.