The latest articles on DEV Community by Konstantin (@konstantin_0e4c2bb99bde9d). https://dev.to/konstantin_0e4c2bb99bde9d https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3257583%2Ff8b6ad9b-5f4a-462a-8928-1737ba81701f.jpg https://dev.to/konstantin_0e4c2bb99bde9d en Konstantin Fri, 12 Jun 2026 06:19:05 +0000 https://dev.to/konstantin_0e4c2bb99bde9d/mcp-of-borisov-geometry-9pn https://dev.to/konstantin_0e4c2bb99bde9d/mcp-of-borisov-geometry-9pn <p><a href="https://github.com/UserPC-SPB/Opterium/tree/v3.1-clean/geoformer" rel="noopener noreferrer">https://github.com/UserPC-SPB/Opterium/tree/v3.1-clean/geoformer</a></p> <p>| 1 | <strong>MCP-native</strong> | Only math MCP server with 71 tools. One config line to integrate with LM Studio/Cline/Claude |<br> | 2 | <strong>Non-standard arithmetic</strong> | Pythagorean table with witness — unique math not found in any other tool |<br> | 3 | <strong>E8-algebra</strong> | Full work with 240 E8 roots: Weyl group, duality, spectrum, triangle geometry — MCP exclusive |<br> | 4 | <strong>Zero-dependency</strong> | Single .exe file. No Python, pip, venv, CUDA. Works on any Windows without installation |<br> | 5 | <strong>Instant start</strong> | Cold start ~289 ms. Persistent mode: requests in microseconds (~50 us) |<br> | 6 | <strong>CubFloat</strong> | Fixed-point arithmetic — no IEEE 754 rounding errors (0.1 + 0.2 = 0.3) |<br> | 7 | <strong>CubComplex</strong> | Exact complex arithmetic: add, sub, mul, pow, conjugate, abs, neg |<br> | 8 | <strong>Vector (19)</strong> | Full vector operations: arithmetic, stats, normalize, clip, sort, unique |<br> | 9 | <strong>Spatial (6)</strong> | 3D space: placement, movement, gravity alignment, distance calculations |<br> | 10 | <strong>Doctor</strong> | Factorization, divisors, primality testing — number diagnostics |<br> | 11 | <strong>Deterministic random</strong> | Seed-based RNG — reproducible results |<br> | 12 | <strong>Multilingual</strong> | Help in 4 languages: EN, RU, ZH, DE |<br> | 13 | <strong>Offline/privacy</strong> | All computation local. No data sent to cloud |<br> | 14 | <strong>Throughput ~14K req/s</strong> | Persistent mode: ~0.05ms per request, 10K+ requests per second |</p> Konstantin Wed, 11 Jun 2025 10:31:15 +0000 https://dev.to/konstantin_0e4c2bb99bde9d/opteriummemory-ram-optimized-ai-memory-for-fast-low-footprint-recall-29mp https://dev.to/konstantin_0e4c2bb99bde9d/opteriummemory-ram-optimized-ai-memory-for-fast-low-footprint-recall-29mp <p><a href="https://github.com/UserPC-SPB/Opterium/tree/v3.1-clean/geoformer" rel="noopener noreferrer">https://github.com/UserPC-SPB/Opterium/tree/v3.1-clean/geoformer</a> # Cube v5 — FINAL BENCHMARK &amp; TEST REPORT</p> <p><strong>Binary:</strong> <code>cube_v5.exe</code> (22.82 MB)<br><br> <strong>Platform:</strong> Windows 10, Python 3.11 </p> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Metric</th> <th>Value</th> </tr> </thead> <tbody> <tr> <td>Total tests</td> <td>118</td> </tr> <tr> <td>Passed</td> <td> <strong>115/118</strong> (97.5%)</td> </tr> <tr> <td>Cold start (median)</td> <td><strong>289.4 ms</strong></td> </tr> <tr> <td>Persistent latency (median)</td> <td><strong>~50-70 us</strong></td> </tr> <tr> <td>Throughput (peak)</td> <td><strong>15,882 req/s</strong></td> </tr> <tr> <td>Tools</td> <td>71</td> </tr> <tr> <td>Binary size</td> <td>22.82 MB</td> </tr> </tbody> </table></div> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Tool</th> <th>Latency (ms)</th> <th>Status</th> </tr> </thead> <tbody> <tr> <td>tools/list</td> <td>392.5</td> <td>71 tools</td> </tr> <tr> <td>cubint_add(3,4)</td> <td>286.7</td> <td>result=7</td> </tr> <tr> <td>cubint_mul(7,8)</td> <td>286.7</td> <td>result=56</td> </tr> <tr> <td>cubfloat_add(0.1,0.2)</td> <td>~290</td> <td>0.3</td> </tr> <tr> <td>cubcomplex_add(1+2i,3+4i)</td> <td>287.8</td> <td>re=4,im=6</td> </tr> <tr> <td>e8(1)</td> <td>298.8</td> <td>8D vector</td> </tr> <tr> <td>e8_get_root(0)</td> <td>300.0</td> <td>[2,2,0,0,0,0,0,0]</td> </tr> <tr> <td>e8_duality_check(1)</td> <td>286.3</td> <td>56 partners</td> </tr> <tr> <td>vec_sum([1,2,3])</td> <td>318.7</td> <td>ok</td> </tr> <tr> <td>doctor(12)</td> <td>306.6</td> <td>factors</td> </tr> <tr> <td>help(en)</td> <td>291.5</td> <td>&gt;200 chars</td> </tr> </tbody> </table></div> <p><strong>Median cold start: ~289 ms</strong></p> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Category</th> <th>Avg</th> <th>Median</th> <th>P95</th> <th>Min</th> <th>Max</th> <th>N</th> </tr> </thead> <tbody> <tr> <td>cubfloat</td> <td>0.070ms</td> <td><strong>0.067ms</strong></td> <td>0.090ms</td> <td>0.044ms</td> <td>0.090ms</td> <td>6</td> </tr> <tr> <td>cubcomplex</td> <td>-</td> <td><strong>0.046ms</strong></td> <td>-</td> <td>0.046ms</td> <td>-</td> <td>7</td> </tr> <tr> <td>vector</td> <td>-</td> <td><strong>0.044ms</strong></td> <td>-</td> <td>0.042ms</td> <td>-</td> <td>19</td> </tr> <tr> <td>spatial</td> <td>-</td> <td><strong>0.043ms</strong></td> <td>-</td> <td>0.042ms</td> <td>-</td> <td>6</td> </tr> <tr> <td>e8</td> <td>-</td> <td><strong>0.074ms</strong></td> <td>-</td> <td>0.042ms</td> <td>-</td> <td>15</td> </tr> <tr> <td>cubint</td> <td>-</td> <td><strong>0.119ms</strong></td> <td>-</td> <td>0.068ms</td> <td>-</td> <td>23</td> </tr> <tr> <td>doctor</td> <td>-</td> <td><strong>0.086ms</strong></td> <td>-</td> <td>-</td> <td>-</td> <td>-</td> </tr> </tbody> </table></div> <p><strong>Persistent mode: 43-119 us</strong> — 3 orders of magnitude faster than cold start.</p> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Test</th> <th>N</th> <th>Total (ms)</th> <th>Avg (ms)</th> <th><strong>Req/s</strong></th> <th>OK</th> </tr> </thead> <tbody> <tr> <td>cubint_mul x50</td> <td>50</td> <td>3.1</td> <td>0.050</td> <td><strong>15,882</strong></td> <td>50/50</td> </tr> <tr> <td>cubint_add x200</td> <td>200</td> <td>14.3</td> <td>0.059</td> <td><strong>13,935</strong></td> <td>200/200</td> </tr> <tr> <td>e8 x50</td> <td>50</td> <td>3.6</td> <td>0.059</td> <td><strong>13,809</strong></td> <td>50/50</td> </tr> <tr> <td>vec_sum x100</td> <td>100</td> <td>7.9</td> <td>0.059</td> <td><strong>12,658</strong></td> <td>100/100</td> </tr> <tr> <td>e8_batch x50</td> <td>50</td> <td>5.0</td> <td>0.080</td> <td><strong>9,989</strong></td> <td>50/50</td> </tr> <tr> <td>doctor x100</td> <td>100</td> <td>10.4</td> <td>0.088</td> <td><strong>9,627</strong></td> <td>100/100</td> </tr> </tbody> </table></div> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Feature</th> <th><strong>Cube v5</strong></th> <th>Wolfram Alpha</th> <th>SymPy</th> <th>SageMath</th> <th>NumPy</th> <th>mpmath</th> </tr> </thead> <tbody> <tr> <td>Delivery</td> <td><strong>.exe (1 file)</strong></td> <td>API (cloud)</td> <td>pip (lib)</td> <td>pip (lib)</td> <td>pip (lib)</td> <td>pip (lib)</td> </tr> <tr> <td>Size</td> <td><strong>~23 MB</strong></td> <td>-</td> <td>~50 MB</td> <td>~1-2 GB</td> <td>~30 MB</td> <td>~5 MB</td> </tr> <tr> <td>MCP protocol</td> <td><strong>YES stdio</strong></td> <td>NO</td> <td>NO</td> <td>NO</td> <td>NO</td> <td>NO</td> </tr> <tr> <td>Tools</td> <td><strong>71</strong></td> <td>10,000+</td> <td>200+</td> <td>500+</td> <td>100+</td> <td>50+</td> </tr> <tr> <td>Non-standard arith</td> <td><strong>YES (witness)</strong></td> <td>NO</td> <td>NO</td> <td>NO</td> <td>NO</td> <td>NO</td> </tr> <tr> <td>E8-algebra (240)</td> <td><strong>YES full</strong></td> <td>partial</td> <td>NO</td> <td>YES</td> <td>NO</td> <td>NO</td> </tr> <tr> <td>Float precision</td> <td><strong>Fixed-point</strong></td> <td>Arbitrary</td> <td>Rational</td> <td>Arbitrary</td> <td>IEEE 754</td> <td>Arbitrary</td> </tr> <tr> <td>Complex numbers</td> <td>YES</td> <td>YES</td> <td>YES</td> <td>YES</td> <td>YES</td> <td>YES</td> </tr> <tr> <td>Vector ops</td> <td><strong>19 tools</strong></td> <td>YES</td> <td>YES</td> <td>YES</td> <td>YES (opt)</td> <td>YES</td> </tr> <tr> <td>3D space</td> <td><strong>YES (6)</strong></td> <td>YES</td> <td>NO</td> <td>NO</td> <td>NO</td> <td>NO</td> </tr> <tr> <td>Doctor</td> <td><strong>YES</strong></td> <td>NO</td> <td>Partial</td> <td>NO</td> <td>NO</td> <td>NO</td> </tr> <tr> <td>GPU accel</td> <td>NO</td> <td>YES</td> <td>NO</td> <td>YES</td> <td>YES (BLAS)</td> <td>NO</td> </tr> <tr> <td>Symbolic</td> <td>NO</td> <td>YES</td> <td>YES</td> <td>YES</td> <td>NO</td> <td>YES</td> </tr> <tr> <td>Matrices</td> <td>NO</td> <td>YES</td> <td>YES</td> <td>YES</td> <td>YES (BLAS)</td> <td>Partial</td> </tr> <tr> <td>Trigonometry</td> <td>NO</td> <td>YES</td> <td>YES</td> <td>YES</td> <td>YES (ufunc)</td> <td>YES</td> </tr> <tr> <td>LLM integration</td> <td><strong>YES native</strong></td> <td>via API</td> <td>no</td> <td>no</td> <td>no</td> <td>no</td> </tr> <tr> <td>Offline/private</td> <td><strong>YES</strong></td> <td>NO</td> <td>YES</td> <td>YES</td> <td>YES</td> <td>YES</td> </tr> <tr> <td>Price</td> <td><strong>FREE</strong></td> <td>FREE</td> <td>FREE</td> <td>FREE</td> <td>FREE</td> <td>FREE</td> </tr> <tr> <td>Throughput</td> <td><strong>~14K req/s</strong></td> <td>~100 req/s (API)</td> <td>~10K (lib)</td> <td>~1K</td> <td>~100K (BLAS)</td> <td>~5K</td> </tr> </tbody> </table></div> <h3> Unique advantages of Cube v5: </h3> <ol> <li> <strong>Only MCP-native math server</strong> — Wolfram Alpha has no MCP, SymPy/SageMath are libraries</li> <li> <strong>Non-standard arithmetic</strong> — Pythagorean table with witness, unique math without analogues</li> <li> <strong>Zero-dependency</strong> — single exe, no installation required</li> <li> <strong>Offline-first</strong> — all computation local, no internet access</li> </ol> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>#</th> <th>Advantage</th> <th>Description</th> </tr> </thead> <tbody> <tr> <td>1</td> <td><strong>MCP-native</strong></td> <td>Only math MCP server with 71 tools. One config line to integrate with LM Studio/Cline/Claude</td> </tr> <tr> <td>2</td> <td><strong>Non-standard arithmetic</strong></td> <td>Pythagorean table with witness — unique math not found in any other tool</td> </tr> <tr> <td>3</td> <td><strong>E8-algebra</strong></td> <td>Full work with 240 E8 roots: Weyl group, duality, spectrum, triangle geometry — MCP exclusive</td> </tr> <tr> <td>4</td> <td><strong>Zero-dependency</strong></td> <td>Single .exe file. No Python, pip, venv, CUDA. Works on any Windows without installation</td> </tr> <tr> <td>5</td> <td><strong>Instant start</strong></td> <td>Cold start ~289 ms. Persistent mode: requests in microseconds (~50 us)</td> </tr> <tr> <td>6</td> <td><strong>CubFloat</strong></td> <td>Fixed-point arithmetic — no IEEE 754 rounding errors (0.1 + 0.2 = 0.3)</td> </tr> <tr> <td>7</td> <td><strong>CubComplex</strong></td> <td>Exact complex arithmetic: add, sub, mul, pow, conjugate, abs, neg</td> </tr> <tr> <td>8</td> <td><strong>Vector (19)</strong></td> <td>Full vector operations: arithmetic, stats, normalize, clip, sort, unique</td> </tr> <tr> <td>9</td> <td><strong>Spatial (6)</strong></td> <td>3D space: placement, movement, gravity alignment, distance calculations</td> </tr> <tr> <td>10</td> <td><strong>Doctor</strong></td> <td>Factorization, divisors, primality testing — number diagnostics</td> </tr> <tr> <td>11</td> <td><strong>Deterministic random</strong></td> <td>Seed-based RNG — reproducible results</td> </tr> <tr> <td>12</td> <td><strong>Multilingual</strong></td> <td>Help in 4 languages: EN, RU, ZH, DE</td> </tr> <tr> <td>13</td> <td><strong>Offline/privacy</strong></td> <td>All computation local. No data sent to cloud</td> </tr> <tr> <td>14</td> <td><strong>Throughput ~14K req/s</strong></td> <td>Persistent mode: ~0.05ms per request, 10K+ requests per second</td> </tr> </tbody> </table></div> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>#</th> <th>Drawback</th> <th>Description</th> <th>Impact</th> </tr> </thead> <tbody> <tr> <td>1</td> <td><strong>No GPU</strong></td> <td>All computation on CPU. No CUDA/Metal</td> <td>Medium for E8 batch 10K+</td> </tr> <tr> <td>2</td> <td><strong>N &lt;= 100 for witness</strong></td> <td>Domain limitation for witness arithmetic</td> <td>Limits scale</td> </tr> <tr> <td>3</td> <td><strong>No symbolic computation</strong></td> <td>Numerical only. No symbolic differentiation/integration</td> <td>No calculus</td> </tr> <tr> <td>4</td> <td><strong>No matrix operations</strong></td> <td>No det, eig, LU, QR, matrix multiplication</td> <td>Key gap for science</td> </tr> <tr> <td>5</td> <td><strong>No trigonometry</strong></td> <td>sin, cos, tan, atan missing</td> <td>Limits physics/engineering</td> </tr> <tr> <td>6</td> <td><strong>No stat distributions</strong></td> <td>No normal, Poisson, chi-squared</td> <td>Basic stats only</td> </tr> <tr> <td>7</td> <td><strong>Single-threaded</strong></td> <td>One thread. Doesn't utilize all CPU cores</td> <td></td> </tr> <tr> <td>8</td> <td><strong>Windows-only</strong></td> <td>.exe file — no Linux/macOS version</td> <td></td> </tr> <tr> <td>9</td> <td><strong>No PyPI package</strong></td> <td>No pip install. Manual download required</td> <td></td> </tr> <tr> <td>10</td> <td><strong>Closed-source format</strong></td> <td>Binary cannot be modified without rebuild</td> <td></td> </tr> <tr> <td>11</td> <td><strong>RSA factorization, matrix multiplications, quantum geometry</strong></td> <td>are removed for security against criminal and military use</td> <td></td> </tr> </tbody> </table></div> <h3> Segment 2: Agent Systems (Cline, Claude Code, Codex) 4/5 </h3> <ul> <li>LLM agents <strong>need accurate computation</strong> — LLMs compute poorly</li> <li>Cube v5 provides 71 tools without external APIs or internet</li> <li>Built-in diagnostic tools (doctor, witness) help LLMs understand numbers</li> </ul> <h3> Segment 3: Education / Science 3/5 </h3> <ul> <li>Group theory, E8-algebra, Pythagorean table — educational topics</li> <li>Doctor diagnostics — good teaching tool</li> <li>But: no symbolic computation, loses to SymPy/SageMath for education</li> </ul> <h3> Segment 4: Finance 3/5 </h3> <ul> <li>Fixed-point arithmetic useful for financial computations</li> <li>Witness structure can be used for pattern analysis</li> <li>Vector statistics for time series</li> </ul> <p><strong>Cube v5 is a unique MCP server</strong> with 71 math tools, having no direct analogues in the MCP ecosystem.</p> <p>Key strengths:</p> <ul> <li> <strong>First free math MCP server</strong> — no competitors in niche</li> <li> <strong>Non-standard arithmetic</strong> — completely unique math model</li> <li> <strong>E8-algebra</strong> — exclusive functionality for LLM tools</li> <li> <strong>~14,000 req/s</strong> in persistent mode — sufficient for real-time</li> <li> <strong>~289 ms cold start</strong> — fast startup</li> <li> <strong>22.82 MB</strong> — compact standalone exe with no dependencies <strong>Key opportunities:</strong> Growing MCP market, LLM agent integration, educational potential.</li> </ul>