Research, case studies, and engineering deep-dives from the HeyNEO team.
NEO built a YAML-driven policy enforcement framework for AI agents with AST-restricted conditions, staged PII detection, JSONL audits, and a FastAPI + React dashboard so rule violations are visible and enforceable.
NEO built a post-hoc context debugger that reconstructs the exact context window at any turn, tracks when facts drop out, and finds divergence points across two agent runs.
NEO built a transparent OpenAI, Anthropic, and Ollama proxy that streams context composition, token usage, evictions, and attention density to a live dashboard.
NEO built a git-backed CLI for AI team configuration sync with tree-level three-way merge, conflict-safe pull and push flows, and a 52-point security and compliance audit.
NEO built an ASR benchmarking harness that compares five speech models across 15+ scenarios with WER, CER, RTF, and inference-time outputs in a stable JSON schema.
A full case study on replacing gut-feel RAG tuning with LLM-as-judge evaluation, retrieval fixes, and model sweep benchmarking that delivered +19% quality and -79% session cost.
A complete evaluation of qwen3.6:27b, qwen3.6:35b-a3b, qwen3-coder:30b, and deepseek-coder:33b across code generation, function calling, and agent capabilities — run entirely on local hardware.
NEO built Agent Factory GUI, a drag-and-drop visual environment for composing, testing, and exporting production AI agent pipelines without writing boilerplate.
NEO built a pipeline that converts arxiv research papers into podcast-style audio episodes with two AI hosts, TTS voice synthesis, background music, and MP3 output.
NEO built AutoDoc, an autonomous agent that traverses your codebase, understands structure and intent, and generates accurate, up-to-date documentation without manual effort.
NEO built a multimodal RAG with CLIP embeddings and ChromaDB: 0.030s retrieval, 60%+ cross-modal accuracy, ingestion for PDFs, images (OCR), and tables (schema extraction).
NEO built a PyTorch pipeline that schedules training around grid carbon intensity and tracks emissions with CodeCarbon—43.2% CO2 reduction on MNIST with accuracy within 0.3% of baseline.
