Back to Projects
live2025 - Present
DeepCrew AI
Personal
deepcrew-ai is a PyPI library (v0.2.0) that provides a clean abstraction layer over the raw primitives every multi-agent Python app needs: provider-agnostic LLM calls via litellm (100+ providers, one API), automatic JSON Schema generation from @tool-decorated functions, parallel DAG workflow execution, and typed real-time streaming. v0.2.0 adds APEX — a confidence-scoring synthesizer with source citation — alongside auto agent spawning with intelligent tool allocation, looping methodology, a Skills system, pluggable memory providers (in-memory and file-backed), per-agent retry and fallback chains, OpenTelemetry observability, and a YAML-driven CLI. Every feature is additive and v0.1.0 programs run unchanged.
Tech Stack
PythonlitellmPyPILLMMulti-AgentAgentic AIOpenTelemetryPydanticasyncio
Highlights
- APEX synthesizer with confidence scoring and source citation across 100+ LLM providers
- Auto agent spawning — any agent can create sub-agents mid-run with intelligent tool allocation
- Production reliability via per-agent retry policies and model fallback chains
- OpenTelemetry observability: spans for every LLM call, tool execution, and workflow step
- Pluggable memory providers (in-memory, file-backed, custom ABC) for stateful agents
- CLI and YAML workflow runner for codeless pipeline execution
Challenges
- Designing a retry system that works with Python coroutines — coroutines can't be restarted, so the retry wrapper takes a factory function instead of a pre-built coroutine
- Making APEX confidence scoring reliable without a dedicated fine-tuned model — uses structured LLM output with JSON response_format and a second synthesis pass
- Wiring OpenTelemetry spans with zero overhead when disabled — all span context managers return nullcontext() when observability is not configured