Built and open-sourced an autonomous local coding agent powered by Google Gemini 2.0 with deep reasoning: B1.

Most AI coding setups run into two walls: recurring API bills on large context windows, or truncated reasoning tokens. B1 takes a different approach:

- Reverse-engineered Web2API proxy running locally on port 8081 (zero API keys required, full reasoning tokens).

- Local Flask backend on port 5000 with a 10-step ReAct agentic loop and Self-RAG query decomposition.

- 6 native system tools for direct file read/write, AST graph search, and PowerShell command execution.

- Linear/Apple HIG glassmorphic interface with interactive live HTML/Canvas artifacts and voice synthesis.

- Automated single-command diagnostic runner python check_health.py) and a 6-phase startup SOP for new clones.

Code and architecture docs are open on GitHub:

https://github.com/IshaanYK/gemini-1b-local-agent

#OpenSource #AI #Python #SoftwareEngineering #Gemini

Built and open-sourced an autonomous local coding agent powered by Google Gemini 2.0 with deep reasoning: B1.

Most AI coding setups run into two walls: recurring API bills on large context windows, or truncated reasoning tokens. B1 takes a different approach:

- Reverse-engineered Web2API proxy running locally on port 8081 (zero API keys required, full reasoning tokens).

- Local Flask backend on port 5000 with a 10-step ReAct agentic loop and Self-RAG query decomposition.

- 6 native system tools for direct file read/write, AST graph search, and PowerShell command execution.

- Linear/Apple HIG glassmorphic interface with interactive live HTML/Canvas artifacts and voice synthesis.

- Automated single-command diagnostic runner python check_health.py) and a 6-phase startup SOP for new clones.

Code and architecture docs are open on GitHub:

https://github.com/IshaanYK/gemini-1b-local-agent

#OpenSource #AI #Python #SoftwareEngineering #Gemini

Fixed a critical model-routing flaw in our open-source local AI agent architecture today.

When reverse-engineering web-to-API proxy layers, it's easy to inherit placeholder model aliases (like fictitious "Gemini 3.8" labels) that route straight into strict Google One Pro endpoints and hang when unauthenticated.

Here is what we rebuilt in Gemini 1B Local Agent:

1. Re-anchored our model registry to official Google models: Gemini 2.0 Flash as the default workhorse (< 350ms TTFT), 2.0 Flash Thinking for transparent Chain-of-Thought, and 2.0 Pro Exp.

2. Built a self-healing proxy fallback: If a user selects a Pro model without session cookies in their .env, the engine automatically catches the routing failure and cascades to high-reasoning Thinking mode—zero dropped requests.

3. Clean SQLite profile lifecycle: Added atomic reset endpoints to wipe vector facts, long-term memory, and session trees for fresh starts.

All pushed and live on GitHub: https://github.com/IshaanYK/gemini-1b-local-agent

#AI #OpenSource #SoftwareEngineering #Python #Gemini #BuildInPublic

Built a zero-friction local AI pair programmer that bridges web reverse proxies and official frontier models.

Many open-source agent setups fail during onboarding due to fragile startup scripts, strict API key requirements, or silent fallback to basic models.

In our latest release of gemini-1b-local-agent:

1. Native OpenAI-to-Gemini Tool Bridge: Enabled autonomous tool calling through standard OpenAI schemas, automatically parsing function calls and executing local OS/workspace operations.

2. Frontier Model Routing: Added direct support for gemini-2.0-pro-exp and gemini-2.0-flash-thinking alongside the local web proxy, allowing developers to switch between free zero-key proxy execution and official Google AI Studio endpoints.

3. Frictionless Onboarding: Added automated environment setup, self-healing socket retries, and comprehensive architecture documentation.

Code & Architecture: https://github.com/IshaanYK/gemini-1b-local-agent

Most people struggle to run local AI agents because of high API costs or brittle setup steps.

I just pushed an update to our open-source B1 Gemini Local Agent repo to solve this:

1. Reverse-Engineered Web2API Proxy: Bridges Google Gemini's raw web RPC into a 100% OpenAI-compatible endpoint (/v1/chat/completions) with zero API keys required.

2. Native Tool Calling Bridge: Translates standard OpenAI tools definitions into structured model instructions, intercepts tool_call blocks with regex parsers, and feeds execution results back into multi-turn ReAct loops (run_command, read_file, write_file, list_dir).

3. Zero-Friction Setup: Cut startup hang from 15 seconds to under 1 second by optimizing build discovery, raised request timeouts to 120s, and automated local permission grants.

The full architectural breakdown is now live in HOW_IT_WORKS.md: https://github.com/IshaanYK/gemini-1b-local-agent

What patterns are you using to bridge web models to local agentic tool calling?

Most AI voice assistants fail for two reasons: high latency from cascading API hops, and fragile turn-taking that freezes after the first reply.

Over the past few weeks, I’ve been building an open-source, fully local AI agent with a native full-duplex Speech-to-Speech (STS) pipeline. Here is the complete architecture breakdown of what’s running under the hood:

1. Unified Speech-to-Speech (STS) Pipeline

Instead of the standard 3-hop cascade (Client STT → LLM generation → regex sentence chunking → separate TTS API calls), everything runs through a single /api/voice/sts endpoint. Spoken input returns a direct, high-fidelity neural audio stream (Edge-TTS en-US-JennyNeural) in one unified hop.

2. Sub-10ms Conversational Reflex Engine

To hit instant human conversational timing, common greetings, confirmations, and interruptions bypass LLM inference entirely. A local reflex matcher resolves and caches replies in under 10ms with natural human conversational fillers ("Oh hey!", "Right!").

3. Zero-Delay Reasoning on Gemini 3.8 Flash

For novel reasoning queries, the pipeline calls Gemini 3.8 Flash with zero reasoning overhead (think: 4, 85-token cap). It delivers punchy, spoken-first answers with natural disfluencies in under 1.5 seconds.

4. Solving Chromium's WebSpeech Deadlock Bug

In browser voice interfaces, calling recognition.abort() tears down Chromium's WebSpeech engine, throwing InvalidStateError on subsequent turns and leaving the mic frozen. I replaced it with a clean instance lifecycle and an autonomous 1200ms duplex watchdog that keeps the listener armed across continuous back-to-back turns.

5. Local Self-RAG Vector Memory

Continuous conversation history is embedded locally on CPU using all-MiniLM-L6-v2. The agent recalls past project context, code snippets, and preferences across sessions without context window bloat.

6. Autonomous Agent Core & Self-Healing Proxy

Backed by 13 active MCP servers and 39 tools across local filesystem execution and code inspection, supported by a self-healing background proxy daemon on port 8081.

Open source code and architecture:

https://github.com/IshaanYK/gemini-1b-local-agent

#AI #MachineLearning #SpeechAI #VoiceAI #Python #SoftwareEngineering #OpenSource #FullStack #BuildInPublic