End-to-end AI systems, data pipelines, and software platforms — from production deployments to research prototypes.
Open-source LLM runtime that abstracts OpenAI, Anthropic, and Gemini behind a unified async client with streaming, structured output validation, retries, circuit breakers, and provider routing. Powers six production AI products including CanApply's Dana platform.
Research-stage learnable scoring layer that separates access frequency from semantic salience in LLM memory systems — the failure mode that causes drift and recall failures in Mem0, Letta, and Zep on long-horizon tasks.
Production DAG-based agent orchestration runtime with 41 domain operators, 20+ tools, SSE streaming, multi-provider LLM execution, JSON Schema validation, and human-in-the-loop gates — serving CanApply's educational guidance workflows.
AI-native B2B SaaS platform that replaces manual resume screening and early-stage interviews with automated evaluation pipelines, LLM-based interview analysis, and candidate ranking systems.
End-to-end news intelligence pipeline ingesting 23 sources with OpenAI-powered summarization, Elasticsearch semantic search, REST APIs, and Telegram delivery.
Production FastAPI backend combining multi-layer rule filtering with a structured 22-question rubric evaluated by an LLM agent for compatibility scoring. Powering 100+ monthly bookings.
End-to-end BTC/USDT futures forecasting pipeline combining Amazon Chronos-2, LightGBM quantile models, regime-aware strategy logic, and walk-forward backtesting.
Graph neural model predicting drug interaction severity using SMILES-based molecular graphs and RDKit descriptors. Benchmarked GAT, GIN, and MPNN variants across 16,000+ structures.
Full-stack monorepo combining Next.js public site, FastAPI backend, MDX + structured content, Clerk-protected dashboard, MongoDB persistence, Telegram command ingestion, LaTeX-based PDF export, and bounded LLM-assisted resume tailoring.
Gym-compatible blackjack environment with partial observability and legality-aware action masking. PPO and DQN agents trained via Stable-Baselines3; DQN achieved a 49.4% loss rate, outperforming the house edge.