whoami open source recently shipped how I build agents experience stack activity
Intake figures come from the production medical agent I led at ReinHealth. Test counts come from CORTEX's unit suite and live test battery.
Local-first by default. Most of this runs entirely on hardware you own, with no API keys.
How CORTEX works: the architecture, the guardrails, and a quick start
flowchart LR
UI["Next.js UI"] -->|SSE| API["FastAPI gateway"]
API --> K["Agent kernel<br/>plan · act · reflect"]
K -. hard tasks .-> LATS["LATS tree search"]
K -. big jobs .-> SUP["Supervisor → workers"]
K --> MEM["Memory<br/>working · episodic · semantic · procedural"]
K --> TOOLS["Sandboxed tools<br/>python · files · web · docs"]
K --> R["Model router"] --> O["Ollama"]
API -. OTLP spans .-> J["Jaeger"]
Why it behaves on a 7B model. Small models break rules that prompts ask them to follow, so CORTEX enforces the rules in code:
- Plans that need no tools run with no tool schemas at all, and destructive plans become refusals before anything executes.
- Pasted or quoted text is treated as data, so instructions hidden inside it can't trigger a tool.
- Web fetch reaches public hosts only. Loopback, private-network and cloud-metadata addresses are refused after DNS resolution and on every redirect.
- Memory stores only durable facts you state about yourself.
git clone https://github.com/roydonsequeira/CORTEX-Private-Intelligence-Framework.git && cd CORTEX-Private-Intelligence-Framework
ollama pull qwen2.5:7b && ollama pull nomic-embed-text
pip install -e . && cortex doctor && cortex serve # API on :8000
cd ui && npm install && npm run dev # UI on :3000
| Project | What it is | Last push |
|---|---|---|
| CORTEX-Private-Intelligence-Framework Python · ★ 4 |
Private, local-first AI agent: planning, sandboxed tools, four-tier memory, streaming UI and OpenTelemetry — runs entirely on your machine with Ollama. | 2026-10-07 |
| Skin-Lesion-Segmentation-in-TensorFlow-2.0 Python |
Skin lesion segmentation on ISIC 2018 using U-Net (PyTorch) and ResU-Net (TensorFlow 2). Final year project. | 2026-05-25 |
| RagChatbot Python |
Production-quality RAG chatbot: Ollama LLM, ChromaDB vector store, OCR/PDF ingestion, Next.js UI | 2026-04-09 |
| clinicalNote-SOAP | n8n workflows for AI-powered Clinical SOAP note generation (TTT, STT, TTS) | 2026-03-25 |
Refreshed daily from the GitHub API by readme-sync. New public repositories show up here on their own.
flowchart LR
IN["User or trigger"] --> ORCH["Orchestrator"]
ORCH --> PLAN["Plan and reason"]
ORCH --> RET["Retrieve context"]
ORCH --> TOOLS["Call tools"]
RET --> VEC["Qdrant / ChromaDB"]
TOOLS --> SYS["APIs / databases"]
PLAN --> CHECK{"Validate and guardrail"}
VEC --> CHECK
SYS --> CHECK
CHECK -->|pass| OUT["Answer or action"]
CHECK -->|fail| ESC["Fallback or human handoff"]
| Principle | What it looks like in production |
|---|---|
| Rules live in code, not prompts | CORTEX turns destructive plans into refusals before any tool runs, and pasted text can never trigger an action. |
| Keep the data home | Local inference on Ollama: the ReinHealth intake agent sent zero patient records to external APIs. |
| Ground every answer | Retrieval over Qdrant, ChromaDB and PostgreSQL instead of trusting model memory. |
| Escalate instead of guessing | Emergency-symptom escalation, input validation and audit logging in the clinical agent. |
| Trace everything | OpenTelemetry spans across model, tool and memory calls, so a failure is debugged as a system rather than guessed at as a prompt. |
AI Agent Developer · Code Crew Studio · Feb 2026 – present · Mumbai (remote)
- Built and maintain a production onboarding agent on the official WhatsApp Business API that runs structured intake conversations, classifies user problems and returns analyzed feedback.
- Engineered a multi-turn query system with persistent context across sessions, using LangChain for orchestration-heavy flows and direct LLM API calls where latency matters.
- Built FastAPI services on PostgreSQL for agent state, conversation history and structured customer records.
GenAI Engineer · ReinHealth, stealth AI healthcare startup · Jul 2024 – Feb 2026 · Colorado, US (remote)
- Led end-to-end delivery of a production autonomous medical intake agent over text and voice, cutting manual intake work by an estimated 70% and average intake time from 15 minutes to under 5.
- Architected an LLM orchestration layer on local Ollama inference (intent analysis, follow-up selection, tool execution, grounded synthesis), so zero patient records reached external APIs.
- Built a privacy-first RAG pipeline on Qdrant with PostgreSQL, and automated speech-to-text clinical notes, TTS summaries and conflict-aware scheduling in n8n.
- Hardened the platform for clinical use with input validation, emergency-symptom escalation and audit logging, behind a single REST API for agents, workflows and frontend.
Machine Learning Intern · Igeeks Technologies · Jun – Jul 2023 · Bengaluru
- Built CNN, AlexNet and MLP image-classification pipelines on custom datasets, with OpenCV preprocessing and hyperparameter tuning.
Education: B.E. in Artificial Intelligence & Machine Learning, NMAM Institute of Technology (2020–2024) · Executive PG Certification in Data Science & AI, iHUB DivyaSampark, IIT Roorkee (2024–2026)
| Layer | Tools |
|---|---|
| Agents & LLMs | LangChain · ReAct · LATS · supervisor–worker orchestration · Ollama · Claude · Gemini · OpenAI · Hugging Face · STT/TTS |
| Retrieval | Qdrant · ChromaDB · BGE and nomic embeddings · semantic search · PostgreSQL · Redis |
| Backend | Python · FastAPI · Flask · Pydantic · SQL · REST · SSE streaming |
| Automation | n8n · WhatsApp Business API · Twilio |
| ML & vision | PyTorch · TensorFlow · scikit-learn · OpenCV · NumPy · Pandas |
| Ship & observe | Docker · Linux · GitHub Actions · OpenTelemetry · Vercel · ruff · mypy |
| Agent UIs | Next.js · TypeScript · Tailwind CSS |
