About
Constitutional AI with signed, replayable decision traces.
Built by Sahil Kadadekar · Solo architect · Sep 2025 – Present
- 1.46M+ Research Measurements
- 55 Technical Reports
Chimera is a constitutional AI enforcement architecture. Every action routes through an embedding-based safety classifier, escalates to multi-model debate when uncertain, and produces cryptographically signed provenance chains with zero-knowledge proofs. The system self-improves: debate outcomes train the alignment encoder through an RLAIF loop.
Action
a chat turn or a tool action
Router
embedding safety classifier; routine queries take the fast path
Debate
multi-model, when the router is uncertain
Provenance
Ed25519-signed chain with zero-knowledge proofs
RLAIF: debate outcomes train the alignment encoder
What This Is
The architecture
A constitutional AI enforcement system spanning Python and Rust. An embedding fast-path router handles routine queries and escalates uncertain ones to a multi-model debate engine with heat-based escalation and three consensus algorithms. The Rust runtime (7 crates) provides Ed25519 provenance chains, BFT consensus, and zero-knowledge proofs for cross-trust-boundary communication. JARVIS is the agent layer — multi-provider chat, voice (Whisper/Piper), semantic memory, tool execution with human-in-the-loop approval, and proactive intelligence.
The research
1,463,000+ measurements across 55 technical reports. Not wall-clock approximations — CUDA event timing with defined hardware profiles and statistical methodology. Covers model loading, ONNX conversion, TensorRT compilation, KV cache optimization, multi-agent coordination, and safety analysis across Ollama, vLLM, and TGI.
The Ecosystem
- 01
Banterpacks
Core monorepo — 6 subsystems: JARVIS (AI gateway), TDD002 (constitutional router), Chimera (debate engine), TDD005 (Rust runtime with ZK proofs + BFT), RLAIF (self-improving alignment), and Authoring (LLM providers).
Python, Rust
- 02
Banterhearts
ML research platform — inference API, benchmarking infrastructure, AutoOpt agent, safety evaluation framework. Source of 1.46M+ measurements across 55 technical reports.
Python
- 03
Chimeraforge
LLM deployment optimizer on PyPI (v0.34.0). Model-agnostic 5-gate capacity planner (VRAM, Quality, Safety, Latency, Cost) — plans any registry / Ollama / HuggingFace model across 22 GPU profiles, and serves the same numbers to AI assistants over MCP.
Python, Rust
- 04
Chimera Multi-Agent
Muse Protocol — 6-agent content pipeline with ClickHouse analytics. Also the observability control plane (OTel, Datadog, DLQ).
Python
- 05
Chimeradroid
Android companion for JARVIS — voice, chat, session handoff, tool approval, mesh networking, offline-first.
C# / Unity
- 06
Echo
Messaging channel adapters — Slack and Discord bridges to JARVIS. Session tracking, device key auth.
Python
- 07
JARVIS Console
Web console for JARVIS — chat with streaming, control room dashboard, cognitive agent ELO, tool catalog, workflow management.
TypeScript / Next.js
- 08
This Site
Public presence. Episodes generated from git commits, research archive, platform documentation.
TypeScript / Next.js
- 09
Project Wyvern
Embodied autonomy plane. Governed mission execution between Chimera control and PX4/ArduPilot — 5-tier authority hierarchy, cryptographic mission replay, OpenAPI 3.1 mission contract. Phase 0 specs complete; SIM-ONLY MVP in progress on PX4 + Gazebo.
Python, Rust, ROS 2
9 repositories · Python, Rust, TypeScript, C#
About This Site
268 episodes were auto-generated from git commits across the nine repositories (per-commit stream archived 2026-06-26; each stream closes with a full retrospective). A multi-agent pipeline (Chimera Multi-Agent) ingested commits and benchmark data, generated roundtable-style commentary with four AI personas, and published to this Next.js site via GitHub + Vercel.
The research archive surfaces 55 technical reports with phase grouping, searchable titles, and ISR with 15-minute revalidation. Every report links to real measurements and defined methodology.