# Jamin — justjammin

I'm Jamin, aka justjammin. I'm based in Chicago, where I build AI platforms, agent tooling and the backend systems they run on. My background includes software and solutions engineering at Cars.com and business sales at Apple. I care about understanding what someone needs and building software that helps.

These days, I work on training infrastructure, agent orchestration, retrieval and LLM evaluation. I want to be able to rerun an experiment, check the results and trust the system when someone uses it. I come at AI platform and forward-deployed engineering from the backend, with an eye on how people will use what I build. Adopt. Adapt. Adept.

- Website: [ORBTL portfolio](https://justjammin.github.io/)
- AKA: justjammin, nami, ORBTL
- Founded: 2018
- Location: Chicago, IL
- Hobbies: DJing, Music Production, Anime, Gaming, Programming, AI, Web Design, Content Creation
- GitHub: https://github.com/justjammin
- LinkedIn: [Jamin's professional profile](https://www.linkedin.com/in/jamin-echols-59b1998a)

## Knows About

AI platforms, LoRA fine-tuning, PyTorch, Ray, Databricks, Prompt engineering, Agent orchestration, MCP, Context engineering, RAG, Hybrid retrieval, Vector databases, LLMOps, TypeScript, Node.js, Next.js, Vue, Tailwind, Python, Go, C#, PHP, Symfony, REST APIs, Flask, CI/CD, Kubernetes, AWS, Marketing, Analytics, Community, Teaching

## Projects

- [Training Platform](/demos/training-platform/): A working 77-intent classifier: 90.86% test macro-F1, calibrated human-review routing, and 11.71 ms CPU p95. Built on reproducible Kaggle training, full-state recovery, Ray/PyTorch execution, and Databricks reporting workflows.
- [Astra](https://astra-os.justjammin.workers.dev): A portable agent harness that takes work from product intent to verified implementation through five explicit gates, with a live review console and human approvals.
- [xoxo-eval](https://xoxo.justjammin.workers.dev): A head-to-head evaluation workbench for coding agents. Compare Claude Code and Codex on isolated tasks, inspect the evidence, and bring in a blind judge.
- [Legible](https://writelegible.com): Writing tracker for authors, designed as a calmer alternative to existing trackers.

## Lab

- Model Regression Detector — active. Recorded Qwen evaluation on a Kaggle T4: a prompt change drops fixture accuracy from 87.5% to 25%, with immutable evidence and a failing regression gate.
- LLM Cost Autopilot — queued. Queue-driven budget controller that measures LLM spend and applies safe routing and batch policies automatically.
- AI Failure Forensics — queued. Trace-first incident workbench for reconstructing why an AI feature failed across requests, tools, and model calls.
- Self-Healing Docs — queued. GitHub automation that detects stale code examples and opens focused documentation repairs with evidence.
- LLM Output Arbitrator — queued. Graph-based arbitration service that compares candidate model outputs and selects the safest useful response.
- Hybrid RAG Pipeline — queued. Production retrieval pipeline combining vector search, BM25, transformer embeddings, and cross-encoder reranking.
- Semantic LLM Cache — queued. Fast request cache that reuses semantically equivalent LLM work while preserving freshness and safety boundaries.
- Guarded Text-to-SQL — queued. Text-to-SQL gateway that validates generated queries against schema, policy, and resource guards before execution.
- Prompt Experiment Platform — queued. Experiment workspace for versioning prompts, replaying datasets, and comparing model outcomes with reproducible evidence.
- Domain LoRA Pipeline — complete. Working BANKING77 classifier with 90.86% test macro-F1, calibrated routing and a local prediction app. Compare TF-IDF with a Kaggle-trained encoder, recover training state, and carry versioned evaluation artifacts into Databricks.
- Observable LLM Gateway — queued. Provider-neutral LLM gateway with request tracing, Redis-backed controls, and operationally useful telemetry.
- AI Feature Flags — queued. Typed feature-control plane for safely rolling out AI behavior by cohort, model, risk, and measured outcome.
- Eval Dataset Generator — queued. Durable workflow that turns source material into reviewed evaluation datasets with traceable provenance.
- Multimodal Doc Processor — queued. Asynchronous document pipeline that extracts text, tables, and images for human-reviewed multimodal knowledge work.
- Agent Orchestration Lab — queued. Inspectable multi-agent runtime for composing workflows, persisting state, and evaluating orchestration decisions.

## For Agents

- [Reading guide](/AGENTS.md)
- [LLM overview](/llms.txt)
- [Site map](/SITEMAP.md)
- [Profile index](/INDEX.md)