Yasin – AI agent development, Python, LLM

An AI memory system designed to transform unstructured conversation history into structured, persistent memory that can be reused across AI interactions. The system processes raw chat data through a multi-stage pipeline to extract observations, synthesize higher-level insights, and construct a user persona that enables more personalized, context-aware AI behavior.

It introduces a layered memory model, moving from raw interactions to clustered insights and ultimately to reusable cognition-level abstractions. The output is both human-readable and machine-readable, enabling seamless integration with downstream systems such as agent frameworks and MCP-compatible environments.

By combining LLM-based synthesis with clustering and provenance tracking, the system enables long-term memory persistence, reduces repeated context reconstruction, and provides a foundation for personalized, stateful AI systems without reliance on continuous external API calls.

• Designed and implemented a 5-stage AI pipeline to transform raw conversation data into structured memory (observations → learnings → cognitions → persona).
• Developed hybrid AI + clustering workflows (LLMs + K-Means) to identify patterns and synthesize higher-level abstractions.
• Built a persona generation layer optimized for direct use in system prompts and agent initialization.
• Implemented provenance tracking across all pipeline stages to maintain traceability from raw data to synthesized outputs.
• Designed a local-first architecture enabling one-time processing with persistent, queryable memory (no recurring API dependency).
• Integrated the system with MCP-compatible servers to enable real-time memory access for agents and developer tools.
• Optimized ingestion and preprocessing with smart truncation and filtering to improve signal quality and reduce noise.
• Structured outputs in both JSON and markdown formats for dual machine and human usability.

An interactive learning platform needed a scalable way to generate and adapt practice content while ensuring alignment with course material, instructor intent, and student progress. Static workflows and manual content creation limited both speed and consistency, especially across diverse subjects and large volumes of content.

An agentic AI system was developed to act as a decision layer within the platform, dynamically determining how to generate, retrieve, and validate content based on context. The system routes requests across structured course data, retrieval pipelines, and generation strategies, ensuring outputs remain grounded in approved materials and tailored to the learner’s needs.

By combining retrieval, transformation, and validation into multi-step workflows, the solution improves content quality, enforces curriculum constraints, and enables continuous optimization through evaluation and feedback loops. It integrates directly into existing product flows, enhancing personalization and learning outcomes without introducing operational overhead.

• Designed and deployed an agentic AI system to generate and adapt practice content based on course context, student progress, and instructor-defined constraints.
• Implemented dynamic routing and query understanding to select appropriate data sources, generation strategies, and validation steps per request.
• Integrated structured content pipelines with LLM workflows, ensuring generation was grounded in approved course materials and filtered by backend-controlled metadata.
• Built multi-step generation flows combining retrieval, transformation, and validation to improve accuracy and pedagogical alignment.
• Enforced guardrails to restrict generation scope (e.g., assigned content only), reducing hallucination risk and maintaining curriculum integrity.
• Developed evaluation pipelines with human-in-the-loop annotation, offline metrics, and A/B testing to measure quality and learning impact.
• Optimized system performance and output consistency through iterative tuning of prompts, retrieval strategies, and feedback loops.
• Collaborated with product and engineering teams to align system behavior with learning outcomes and platform constraints.

An agentic browser-based system to automatically discover, navigate, and extract structured data from local business websites for a kids activity marketplace. The challenge was to convert highly unstructured and inconsistent web content into a normalized, searchable dataset of activity providers and their offerings.

The system uses AI-guided navigation to explore websites beyond static scraping, identifying relevant pages such as program descriptions, class offerings, and business information. Extracted content is then classified, filtered, and post-processed into structured records, enabling downstream use in search, filtering, and recommendation workflows.

By combining intelligent page selection, content extraction, and validation, the solution transforms fragmented web data into a consistent database of businesses, activities, and metadata, while filtering out irrelevant or unsuitable content.

• Designed an agentic browser workflow to autonomously navigate business websites and identify relevant content pages.
• Implemented AI-guided page classification (e.g., business info vs. offerings) to structure extraction logic.
• Built extraction pipelines to capture key attributes such as business details, activity types, and program offerings.
• Developed post-processing and normalization layers to convert raw content into structured, queryable records.
• Implemented filtering logic to exclude irrelevant or unsuitable sources (e.g., adult-only businesses, external booking tools).
• Handled variability in website structures, ensuring robust extraction across different layouts and content formats.
• Integrated LLM-based parsing to transform semi-structured text into consistent JSON outputs.
• Enabled scalable data ingestion for marketplace use cases such as search, discovery, and recommendation.

An AI-powered assistant developed to centralize company knowledge and provide real-time, context-aware support for both internal teams and customer-facing workflows. The system enables sales representatives to quickly access relevant product details, generate tailored responses, and handle objections, while also supporting customer success teams with faster, more reliable issue resolution.

By combining internal data with intelligent retrieval and task-oriented workflows, the solution improves response quality, reduces manual effort, and ensures a more consistent experience across high-volume interactions. It serves as a scalable layer on top of existing business operations, enhancing both efficiency and customer engagement without requiring major workflow changes.

• Designed and deployed an AI assistant for Sales and Customer Success, enabling real-time access to internal knowledge and reducing manual lookup overhead.
• Built end-to-end RAG pipelines (ingestion, embedding, retrieval) to unify fragmented enterprise data into a searchable system.
• Implemented hybrid retrieval strategies to improve response relevance across product, competitive, and support queries.
• Developed agentic workflows with tool routing to handle multi-step queries (e.g., content generation, sales support, customer issue resolution).
• Integrated LLM outputs into business workflows to support faster response generation and more consistent communication.
• Established evaluation and monitoring pipelines to track response quality, retrieval accuracy, and system performance.
• Iteratively improved system behavior using feedback loops from real user interactions and internal teams.
• Collaborated with stakeholders to identify high-impact AI use cases and align solutions with business needs.