DeepSpec
Reference · not authored by us
Speculative decoding research — training, eval, benchmarks.
Training and evaluating draft models for speculative decoding. Multi-GPU training, full benchmark suite.
PythonPyTorchMulti-GPUML Research
Our own builds — open-source tooling, research, and content operations.
Client engagements have their own case studies.
20 projects
Guardrails and evaluation for agents running with real access.
Reference · not authored by us
Speculative decoding research — training, eval, benchmarks.
Training and evaluating draft models for speculative decoding. Multi-GPU training, full benchmark suite.
Evaluation framework for AI systems — quality metrics, A/B testing, regression detection.
A systematic, reproducible way to measure AI quality with golden test sets, precision/recall/faithfulness/relevance metrics, A/B comparison, HTML dashboards, and regression detection.
AI safety middleware — PII detection, injection defense, moderation, validation.
Security middleware between users and AI models. Detects and mitigates PII leakage, prompt injection attacks, toxic content, and length violations with configurable actions.
The layer between your application and the model providers.
AI agent framework — tool use, memory, multi-step orchestration.
Lightweight AI agent framework with built-in tool calling, persistent memory, and multi-step task orchestration. Designed to run locally or at the edge.
Production-grade context engine — hybrid RAG, reranking, token-aware.
Hybrid RAG pipeline for AI agents with cross-encoder reranking, token-aware windowing, and multi-provider embedding support. Indexes any codebase and retrieves the most relevant context.
OpenAI-compatible AI gateway — rate limiting, caching, failover, cost tracking.
Production-grade AI gateway that sits between your app and LLM providers. Adds rate limiting, smart caching, automatic failover, and per-key cost tracking — drop-in compatible with any OpenAI SDK client.
Unified credential mesh — store, inject, rotate API keys.
Single source of truth for all LLM provider API keys. Encrypts at rest, injects at runtime, rotates without restart. Works across OpenCode, CI/CD, and multi-agent setups.
Intelligent LLM cost router — cheapest model that meets quality bar.
Routes LLM requests to the cheapest model that passes a quality threshold. Tracks cost-per-task, benchmarks model families, and auto-downgrades on budget overflow.
AI code archaeology — index sessions into a searchable knowledge base.
Indexes all past AI coding sessions into a local, searchable knowledge base. Find who wrote what, when, and why — across months of work.
Persistent cross-session memory for AI agents — JSONL local-first.
Gives AI coding agents persistent memory across sessions. Stores conversations, decisions, and context in JSONL files — no cloud, no vendor lock.
Command-line tools for shipping, measuring, and driving code.
Standardized benchmark harness for AI coding-agent skills.
Declarative JSON test suites for evaluating AI agent skills in CI. Measures accuracy, speed, and cost per skill. Drop into any pipeline.
Analytics for OpenCode — costs, models, agents, daily trends.
Tracks and reports on AI coding usage: cost per session, model breakdown, agent performance, daily token trends. Beautiful terminal dashboards.
Scans any codebase, auto-generates OpenCode config — ⭐1.
Walks a codebase, maps its structure, and generates optimized OpenCode configuration — skills, context files, agent routing. Zero config needed.
Skill-building curriculum for AI agents — learn, practice, track.
Structured curriculum for training AI coding agents. Progressive skill levels, practice exercises, and progress tracking across sessions.
Drive opencode on remote machines over SSH.
Launch and manage OpenCode sessions on remote machines via SSH. Detached tasks, PowerShell exec, and bidirectional file sync.
Terminal workspace — registry, launcher, session stats.
A terminal workspace for OpenCode projects. Registry of projects, one-key launcher, session history, and resource usage stats.
Surface and entertainment layers built on the same stack.
60FPS browser puzzle game — zero dependencies. ⭐1.
A cosmic merge game with custom physics, spatial partitioning, and particle effects. Runs smooth on mobile.
985 tech outlets across 50 verticals — accessible PWA.
A searchable directory of hand-curated tech news outlets. Full-text search, bookmarks, dark/light themes.
Publications and broadcasts built on the same stack.
Daily scans of 50+ outlets — signal over noise.
A daily intelligence brief scanning 50+ outlets across 8 verticals. Tech, security, and media — every morning.
Weekly intelligence brief — every Friday.
A weekly podcast on technology, security, and media. Cutting through the noise to what matters.
Send a brief. Clear scope, timeline, and fixed price within 48 hours.
Start Your Project