MCP server exposing Dimitres Kisimov's portfolio optimization/analytics engines as tools for any MCP client.
Six tools with complete JSON schemas and detailed descriptions (avg 180 chars, well above 10-char minimum). All parameters typed and constrained. Strong honesty contract via data_note fields and provenance tracking. However, tool names lack action verbs (forecast_demand, optimize_slotting, pack_cartons are noun-heavy; should be predict_demand, optimize_warehouse_slotting). Error handling is structured but lacks recovery guidance (InvalidInputError, EngineUnavailableError caught but responses don't suggest next steps). No tool annotations (readOnlyHint, idempotentHint) visible in schema. Output schemas documented in descriptions but not formally in JSON Schema. Missing parameter descriptions for some nested object fields (e.g., container.length_cm has description, but the container object itself lacks a top-level description of what it represents).
Discount-leakage drill-down from my sales-kpi-analytics repo, on its SYNTHETIC 24-month B2B order dataset (deterministically generated — no real customer data). Leakage = list value minus revenue; the within-policy vs excess split uses an ASSUMED sanctioned-discount ceiling (policy_discount_pct, default 10) because the dataset carries no contract terms. Returns totals plus the top offenders by sales rep and by region.
Forecast weekly SKU demand with the decision-chain engine. History is REAL data (UCI Online Retail II, ~200 tracked SKUs) cleaned by a provenance-tagged pipeline; forecasts are DERIVED model output chosen per Syntetos-Boylan demand class by lowest mean MASE under rolling-origin cross-validation — MASE-honest: if naive wins a class, that is what gets reported. Pass a sku for its forecast, a demand_class (smooth / erratic / intermittent / lumpy) for the class view, or neither for the scoreboard. First call takes ~10 s (ingest); later calls are cached.
Optimize warehouse slotting on the SYNTHETIC seeded warehouse of my logistics-digital-twin repo (deterministic, seed 42 — no real warehouse data). Solves the exact linear-assignment slotting (Hungarian algorithm), and returns demand-weighted pick travel before/after, the re-shuffle move list ranked by daily saving, and a break-even analysis for executing it.
Pack cartons into containers with First-Fit-Decreasing shelf packing from my logistics-digital-twin repo. FFD is a HEURISTIC (bin packing is NP-hard) and no carton rotation is attempted. Give a list of items (dims in cm, weight in kg, optional quantity) to pack your own cartons, or omit items to pack the repo's SYNTHETIC seeded 60-carton dataset. Returns fill rate, container count vs the one-carton-per-container baseline, and the per-container assignment.
Tool names lack action verbs. 'forecast_demand', 'optimize_slotting', 'pack_cartons' are noun-heavy and do not clearly signal the action to LLMs. Should be 'predict_demand', 'optimize_warehouse_slotting', 'pack_items_into_containers'.
Error responses lack recovery guidance. InvalidInputError and EngineUnavailableError are caught and structured, but responses do not suggest next steps (e.g., 'Try calling forecast_demand with a valid SKU from the scoreboard'). LLMs cannot self-correct without actionable hints.
No tool annotations (readOnlyHint, idempotentHint, destructiveHint) visible in schema definitions. All six tools are read-only and idempotent, but this is not formally declared in the tool metadata. Clients cannot infer safety properties without explicit hints.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2025-06-18+ | v2 |
Read-only quality scorecard for one local portfolio repo, via my portfolio-ops audit engine: git state, expected artifacts (README, LICENSE, CI, tests, business case, deliverables), ruff, and optionally pytest (run_tests=true is slow). This audits real local repositories on this machine; it measures and never modifies them.
Solve a capacitated vehicle routing problem (CVRP) on a SYNTHETIC seeded instance from my route-optimizer repo (tiny / n30 / n60 / n100 customers; deterministic committed JSON, grid read as km). OR-Tools guided local search under a fixed solution limit — deterministic, same input gives the same answer — compared against the Clarke-Wright savings heuristic baseline. Returns km, the saving vs baseline, and the per-vehicle routes with loads.
Output schemas not formally documented in JSON Schema format. Descriptions mention return structure (e.g., 'Returns fill rate, container count, per-container assignment') but no explicit outputSchema field in tool definitions. LLMs cannot parse expected response structure.
Nested object parameters lack descriptions at the container level. 'container' parameter in pack_cartons has sub-fields (length_cm, width_cm, etc.) with descriptions, but the container object itself has no description explaining what it represents or when to use defaults.