chunktort

Hi Jeff,

Streaming millions of geological data points to the browser while keeping it interactive is a real performance architecture problem — the kind where how you structure your data pipeline matters as much as the rendering layer.

Relevant: Chatbot Widget SaaS — Next.js 15 + TypeScript + Supabase + pgvector + Shadow DOM widget, real-time WebSocket streaming, full auth + Stripe billing. Built the whole stack as first engineer. GitHub: github.com/ChunkyTortoise/chatbot-widget. Also: AI Dashboard (live at ai-dashboard-rust.vercel.app) — Next.js 15, shadcn/ui, real-time Recharts, Anthropic streaming.

Stack match: TypeScript, Next.js, Supabase, Python, Vercel — AI-native workflow (Claude + Cursor daily). Available within 1 week, 35-40 hrs/week.

Cayman | [email protected] | github.com/ChunkyTortoise

Hi Oscilar team,

Building reliable AI decisioning for fraud and compliance is a different challenge from most LLM work — the output needs to be auditable and the false-positive cost is real money.

Relevant: Jorge Real Estate AI Bots — production FastAPI + Redis + Claude pipeline, 1,700+ tests, GHL webhook integration with rate limiting, HMAC signature verification, and per-contact Redis locks to prevent race conditions. The same reliability patterns apply to risk decisioning. GitHub: github.com/ChunkyTortoise/jorge_real_estate_bots

Stack match: Python, FastAPI, Claude API, Redis, LLM agent systems — 15+ AI projects in production. Available within 1 week, 35-40 hrs/week.

Cayman | [email protected] | github.com/ChunkyTortoise

Hi Littlebird team,

Persistent memory and context-aware agents are the hard part of personal AI — retrieval has to be fast enough to feel invisible and selective enough not to be noise.

Relevant: GraphRAG demo — entity extraction (spaCy + Claude), NetworkX DiGraph, hybrid BM25+RRF retrieval, fact-checking pipeline. Purpose-built for context that needs structure, not just semantic similarity. GitHub: github.com/ChunkyTortoise/graphrag-demo. Also: MCP Server Toolkit (PyPI published, 233 tests) — memory and context tooling for AI systems.

Stack match: Python, PostgreSQL, RAG, LLM APIs (Anthropic), pgvector — available within 1 week, 35-40 hrs/week.

Cayman | [email protected] | github.com/ChunkyTortoise

Hi Mobasi team,

Autonomous AI agents for forensics is a hard problem — managing tool orchestration across 150+ forensic tools while hitting CJIS compliance requirements is exactly the kind of infrastructure challenge most engineers haven't touched.

Relevant: AI Workflow API — FastAPI + ARQ + Redis workflow engine, YAML-driven, 5 node types (trigger/LLM/condition/HTTP/notify), SSE streaming, 148 tests. Similar distributed orchestration pattern to what you're building. GitHub: github.com/ChunkyTortoise/ai-workflow-api

Stack match: Python, TypeScript, Next.js, Docker, Redis — I use Claude as part of my development workflow daily and ship AI-native. Available within 1 week.

Cayman | [email protected] | github.com/ChunkyTortoise

Hi Hearo team,

Turning Reddit noise into structured product signals is exactly the kind of pipeline problem where RAG architecture matters — naive semantic search on raw posts misses context that entity-aware retrieval catches.

Relevant: GraphRAG demo — hybrid BM25+RRF retrieval on unstructured text, NetworkX entity graph, Claude for extraction, pgvector for embeddings. Built specifically to handle noisy real-world text. GitHub: github.com/ChunkyTortoise/graphrag-demo. Also: DocExtract AI (production, live) — async document pipeline, pgvector + Claude + ARQ worker, 234 tests. docextract-api.onrender.com

Stack match: Python, PostgreSQL, pgvector, RAG, LLM APIs (Anthropic/OpenAI) — every item in your stack is something I ship with. Available within 1 week.

Cayman | [email protected] | github.com/ChunkyTortoise

Python/FastAPI agent work is my core stack: I've built a production multi-agent orchestration layer (LangGraph ContentPipeline, 33 tests), a YAML-driven AI Workflow API with typed context pipelines and condition branching (FastAPI + ARQ, 148 tests), and a GraphRAG system with hybrid BM25+graph-fusion retrieval (63 tests). For eval methodology: the Jorge AI bot ships with a production_e2e_tests.py (806 lines, B8 invariant checks on every agent response). I'm US-based but comfortable with EU remote overlap.

[email protected] | github.com/ChunkyTortoise

I've designed and shipped production autonomous agent systems across voice and text channels: TechNova Voice Bot (FastAPI + WebSocket + Deepgram STT/TTS + Claude, real-time VAD pipeline, 26 tests, demo mode for offline eval) and Jorge AI (multi-turn SMS qualification agent, 1,753 tests, live on Render, with rate-limit enforcement, dedup, human-handoff logic, and end-to-end runbooks). Also built a LangGraph multi-agent pipeline (Research→Draft→Review→Publish, 33 tests) and a GraphRAG system with hybrid retrieval. 19 AI/ML certs including DeepLearning.AI specializations in LLMs and MLOps. Strong on eval methodology and edge-case hardening for real-world agent deployment.

[email protected] | github.com/ChunkyTortoise

The "generative AI + deterministic code" framing is exactly the pattern I've been productizing: I built a YAML-driven AI Workflow API (FastAPI + ARQ + Claude + SSE streaming, 148 tests) where each LLM node has explicit typed schemas and condition branches — AI creativity gated by deterministic routing logic. Also shipped a GHL Multi-Vertical Kit (87 tests, 3 production verticals via config swap) and a GraphRAG demo with hybrid BM25+RRF retrieval. 19 AI/ML certifications (Google, DeepLearning.AI, IBM, Microsoft). US resident, no sponsorship needed.

[email protected] | github.com/ChunkyTortoise

Sending this to [email protected] per instructions, but flagging here too. This is an exact stack match: I've shipped Django/Python + Celery/Redis + Postgres + LLM API pipelines in production — most directly the Jorge real-estate AI bot (auditable human-in-the-loop handoff flow, rate-limited Redis queues, 1,753 tests) and an AI Workflow API with YAML-driven node pipelines, ARQ workers, and SSE streaming. Strong on the 0→1, own-it-end-to-end pattern. PT/US Pacific timezone. Contract-to-FT works well.

[email protected] | github.com/ChunkyTortoise

Hi — this stack is my daily driver. I built Jorge, a live FastAPI + Redis + Claude AI real estate qualification bot (1,753 passing tests, deployed on Render) that handles inbound SMS intake, multi-turn qualification, calendar booking, and human-handoff logic via GHL CRM. I also built DocExtract AI (FastAPI + ARQ workers + pgvector + Supabase, live at docextract-api.onrender.com) and an agentic workflow API with SSE streaming and YAML-driven pipelines. The autonomous admin burden elimination framing — scheduling, intake, interoperability — maps exactly to what I've shipped in production.

[email protected] | github.com/ChunkyTortoise

Built this after hitting real latency issues in production. LangChain added 250ms+ overhead before the LLM was even called — routing, validation, and middleware stacked up fast.

The core realization: most "orchestration" is just routing + retry + caching. Those three things don't require 47 packages or 15MB of dependencies.

What it actually is: ~15KB of Python, 2 runtime deps (httpx + pydantic). Benchmarks on 1,000 requests showed 65ms avg vs LangChain's 420ms, 3MB memory/request vs 12MB.

The part I'm most proud of: the testing layer. MockLLMClient lets you inject deterministic responses so you can test agent routing and retry logic without hitting any APIs. Hard to do that cleanly with LangChain's abstractions.

214 tests, 91% coverage. Live demo at ct-agentforge.streamlit.app if you want to try it without cloning.

Happy to go deep on any part of the architecture.