Genrikh: Building a $0-Marginal-Cost Universal AI Problem Solver with Gemma 2, Sentry & ElevenLabs

Submission for the Hacktoberfest Open-Source AI Challenge: Week 2 — Side Hustle.

Repository: github.com/torrua/triz-universal

Author: @torrua

Target Partner Tracks: Best Use of Gemma, Best Use of Sentry Agent Tracing, Best Use of Render, Best Use of ElevenLabs

In 1946, Soviet inventor and patent examiner Genrikh Saulovich Altshuller (1926–1998) began reviewing thousands of breakthrough inventions to answer an audacious question: Is inventive genius a genetic lottery, or is it an exact algorithmic science?

Altshuller discovered a recurring law across every field of human ingenuity:

Real innovation is NOT compromise.

Conventional thinking resolves conflicting requirements by meeting in the middle: buying more hardware to handle unoptimized code, crippling battery life for faster sensor reads, or raising prices to cover cloud overhead.

Inventive thinking sharpens the conflict into an atomic Physical Contradiction and eliminates it completely using latent resources already present in the system.

Altshuller formalized this into TRIZ (Theory of Inventive Problem Solving) and the rigorous ARIZ algorithm.

When modern AI chatbots encounter tough trade-offs, they default to lazy, expensive platitudes: "In the real world, trade-offs are inevitable. You should purchase an enterprise Redis cluster or scale your cloud instances."

I built Genrikh to change that.

Genrikh is an autonomous, open-source AI problem solver that brings classical ARIZ algorithmic rigor to modern engineering, hardware, and business systems—with $0 marginal cost and 100% intellectual property privacy.

When solo-founders, freelance developers, and small engineering teams build products, they hit crippling deadlocks:

Hiring a human senior architect or patent attorney costs $250–$400/hour.

With Genrikh, any freelance engineer or consultant can run an audit in 60 seconds and deliver:

Large Language Models tend to hallucinate or invent post-hoc justifications if asked for an answer directly. To enforce scientific rigor, Genrikh implements a Dual-Loop Execution Pipeline:

flowchart TD
    subgraph Loop1["Loop 1: Deep Methodological Reasoning (Internal State Lock)"]
        S0["Step 0: Limit & Constraint Pre-Check<br/>(Hard vs Soft vs Assumptions)"] --> S1["Step 1: Mini-Problem & IFR Framing<br/>(Zero Added Cost Goal)"]
        S1 --> S2["Step 2: Sharpen Physical Contradiction<br/>(Element X must be P and NOT-P)"]
        S2 --> S3["Step 3: Substance-Field Resource Audit (VPR)<br/>(Latent System Resources)"]
        S3 --> S4["Step 4: Resolution Strategies<br/>(Space, Time, Condition, Structure, Satisfy, Bypass)"]
        S4 --> S5["Step 5: Verification & Secondary Harm Audit<br/>(Empirical Experiment & Metric Plan)"]
    end

    subgraph Loop2["Loop 2: Contextual Output Delivery (Progressive Disclosure)"]
        L1["Layer 1: Plain-Language Executive Core<br/>(Zero Jargon for Clients & Management)"]
        L2["Layer 2: Professional TRIZ Passport<br/>(Auditable Specs for Patent & Tech Leads)"]
        L3["Layer 3: ElevenLabs Audio Debriefing<br/>(Spoken Voice Briefing for Pitches)"]
    end

    Loop1 --> Loop2
    S5 -.->|Locks State| L1
    S5 -.->|Locks State| L2
    S5 -.->|Locks State| L3

For consultants dealing with trade secrets, proprietary schemas, and pre-patent hardware designs, data privacy is non-negotiable. You cannot legally paste client IP into commercial closed-source APIs.

Genrikh uses Google Gemma 2 running locally via Ollama (http://localhost:11434) as its primary intelligence engine:

<start_of_turn>user ... <end_of_turn>). google-genai) when high cloud throughput is desired. AI agents must not be black boxes. Genrikh integrates Sentry Agent Tracing to instrument every phase of the ARIZ pipeline:

genrikh.step0_constraints, genrikh.step1_mini_problem, genrikh.step2_physical_contradiction, genrikh.step3_resources, genrikh.step4_strategies, and genrikh.step5_dual_loop.

with trace_ariz_step("step2_physical_contradiction", {"param_a": param_a, "param_b": param_b}):
    pc = sharpen_physical_contradiction(problem, param_a, param_b)

Deploying the consultant for your team or agency takes 60 seconds with Render:

render.yaml. Dockerfile running the FastAPI dashboard. Consultants often pitch to non-technical stakeholders (founders, CFOs, product managers) who don't have time to read a 10-page technical memo.

genrikh/voice.py)..mp3), delivering the core architectural decision and ROI metrics in under 60 seconds. We created an automated benchmark suite (examples/gemma2_benchmark.py) testing Genrikh across three distinct domains:

Benchmark Score: 100% contradictions resolved without introducing middle-ground trade-offs.

git clone https://github.com/torrua/triz-universal.git
cd triz-universal
pip install -e .
genrikh solve \
  --problem "High database read concurrency locks write transactions" \
  --param-a "Read Concurrency" \
  --param-b "Write Latency & Storage Cost"
genrikh consult --template freelancer_offline_sync --voice --voice-out briefing.mp3
genrikh serve --port 8000

Genrikh also serves as an agent skill bundle for Google Antigravity, Claude Code, and Cursor (skills/triz-universal), including all 16 canonical reference manuals and the complete Altshuller 39x39 contradiction matrix.

If you believe that true engineering means eliminating compromises rather than settling for them, give the project a star on GitHub! ⭐