Kiah Box Projects & Build Notes

Live demos, product R&D, and operator workflows from the stack OttoBot uses in public.

Featured Project

SPEC (Beta): Prompt Compactor + Shorthand Codec

Plain-English prompt compaction • lossless ~-token codec • opt-in integration

The Problem

Large Language Models process tokens sequentially, making long prompts expensive and slow. Current solutions either lose context (summarization) or maintain full size (no optimization).

The Solution

SPEC is an experimental prompt pre-processor designed to reduce prompt tokens and latency (model-dependent), while avoiding breakage on structured content. Output stays usable for any LLM.

  • Token-Focused Compaction - removes low-information filler and stopwords (lossy by design)
  • Protected Segments - leaves code blocks, inline code, URLs, emails, paths, and quoted strings untouched
  • Profiles - safe, balanced, aggressive trade safety vs savings
  • Shorthand Codec - optional lossless encode/decode via ~-tokens (for storage/transport)
  • Go/No-Go Benchmark - uses Ollama prompt_eval_count + latency to measure real token impact

Results

Profiles
Safety vs savings
Protected
Code/URLs safe
Benchmarked
Ollama metrics
Opt-In
Off by default

Example

Before:
"Please summarize this information and generate a short response for the assistant."
↓ SPEC Compression
After:
"summarize information generate short response assistant."
Lossy (prompt-only) Protects code/URLs Measure tokens via Ollama

Status & Roadmap

✅
Phase 1: Core Compactor

Profiles + protected segments + whitespace normalization

✅
Phase 2: Go/No-Go Benchmark

Ollama prompt_eval_count + latency A/B runner

✅
Phase 3: Demo

Interactive web demo with safe/balanced/aggressive profiles

⏳
Phase 4: Quality Gates

Edge-case suite + meaning regression checks + persona/history compaction

⏳
Phase 5: Integration

Opt-in integration after benchmarks + regression checks are consistently green

Technical Details

🐍 Language

Python (core) + JS (demo)

🛡️ Safety

Protected segments + profiles

📏 Benchmarking

Ollama go/no-go metrics

🔌 Integration

Feature-flagged (off by default)

Impact

SPEC helps reduce prompt token waste for local/private LLM workflows (results vary by model/tokenizer):

  • Lower prompt token count and faster prompt evaluation (often the bottleneck)
  • Protects structured content (code/URLs/paths) from accidental corruption
  • Compactor is not reversible (prompt-only); codec is lossless for dictionary tokens
  • Integration stays opt-in until quality gates are met
Try Interactive Demo Try Codec See Tools

Other Active Projects

🤖 Autonomous AI System

Active

Self-improving AI with live modules, operating telemetry, and autonomous workflows running against real websites and internal tasks.

Live proof Local-first Operator stack
Read the build log →

📊 Knowledge Base

Active

Local semantic retrieval across operator docs, references, and structured business knowledge.

Search Reference Vector search
See a live tool →

🔄 Self-Improvement Engine

Active

LLM-powered continuous code analysis, refactoring, and optimization.

Autonomous Git-integrated 24/7 operation
See the SPEC workflow →

💰 Trading Bot

Development

AI-powered trading with LLM signal confirmation, risk management, and paper trading.

Paper trading Multi-strategy Risk managed
See related tools →

📝 Content Generation

Active

Automated blog posts, social media updates, and technical documentation with SEO optimization.

SEO optimized Multi-format Autonomous
View Blog →

🌐 Web Integration

Active

Search APIs, web scraping, and social media posting across multiple platforms.

SerpAPI Tavily Twitter/X
See tools →

Exploring Next

🧠 Multi-Agent Collaboration

Multiple AI agents working together on complex tasks

Research track

🎯 Reinforcement Learning

Self-training through environment feedback

Evaluation track

🔗 Blockchain Integration

On-chain AI operations and smart contracts

Optional expansion