Stack: FastAPI · LangChain · Celery · SQLAlchemy · React · Docker

mail-phis/

SOC-grade phishing detection for raw MIME email. Nine-stage forensic pipeline: header analysis, SPF/DKIM/DMARC verification, URL threat intelligence, domain age, homograph detection, NLP social-engineering scoring, attachment risk, dual-bucket scoring, IOC export.

Raw MIME → 9-stage pipeline → Verdict (SAFE / MARKETING / SUSPICIOUS / PHISHING) + IOC bundle

Key features: dual-bucket scoring (suspicion − trust, 0–100) · content-only NLP capped at SUSPICIOUS · STIX2 IOC export · concurrent OpenPhish / PhishTank / URLhaus lookups · 16 pytest tests (CI passing)

Stack: FastAPI · asyncpg · Celery · Docker

company-builder/

Autonomous company-building experiment: one goal prompt takes an orchestrator agent from open-internet pain research to a finished go-to-market package with no human input mid-run. The run produced ReconStock (a safety-first Shopify inventory-sync concept): verified market research, an idea tournament with adversarial fact-checking, business plan, brand identity, a working product demo + landing page, launch/founder videos, and a red-team report — every claim traced to a fetched URL or disclosed in an honesty log.

Pain hunt (8 parallel researchers) → Tournament + adversarial verification → Business design → Brand → Build + verify → Videos → Red team → Package

Key features: multi-agent orchestration (parallel researchers, judge panels, skeptic verifiers, fresh-eyes auditor) · self-correction (caught its own misattributed stat and synthetic evidence) · 34-check Playwright end-to-end verification · fully local video production (Playwright screen recording + TTS + ffmpeg) · decision + honesty logs · start at company-builder/run-1/RECAP.html

Stack: Claude Code multi-agent orchestration · Playwright · Node · ffmpeg-static

Core pattern

Every agent in this repo runs a ReAct loop — Reason, Act, Observe, repeat. The LangGraph graphs are structured expressions of this loop. The bare-API version makes the mechanics explicit:

messages = [{"role": "user", "content": task}]

while True:
    response = client.messages.create(model=..., tools=tools, messages=messages)
    messages.append({"role": "assistant", "content": response.content})

    if response.stop_reason == "end_turn":
        return final_answer(response)

    tool_results = [run_tool(block) for block in response.content if block.type == "tool_use"]
    messages.append({"role": "user", "content": tool_results})

messages is the agent's memory — every tool call and result is appended so the LLM always sees the full history of what it has done. LangGraph replaces the while loop with a StateGraph, tool dispatch with ToolNode, and the stop condition with an END edge — same loop, more structure.

Stack