Inspiration
Public administrations and municipal governments face an acute, systemic challenge: auditing and depurating unfunded pension liabilities. In territories across Latin America and emerging markets, public entities manage pension debts running into millions of dollars based on decades of fragmented, physical, and non-standardized labor histories.
Auditing a single municipal employee's pension file traditionally requires weeks of legal review, manual cross-referencing against shifting administrative jurisprudence, and complex actuarial recalculations. A single data entry mistake can cause severe fiscal damage or costly litigation. We founded Pensiotech to replace this error-prone, manual bottleneck with an autonomous, agentic audit engine built on the Model Context Protocol (MCP).
What It Does
Pensiotech transforms complex, unstructured public pension records into auditable, legally defensible, and mathematically verified fiscal reports:
- Intelligent Labor Record Ingestion: Agents ingest multi-page scanned administrative resolutions, contribution logs, and employment histories, converting messy documents into structured chronological timelines.
- Regulatory & Legal Verification: An autonomous agent cross-checks employment gaps, overlapping periods, and applicable retirement regimes against administrative statutes and legal precedents.
- Deterministic Actuarial Math: Bypassing LLM calculation drift, agents interface through MCP with formal calculation engines to compute Base Liquidation Income ($IBL$), contribution density, and replacement rates.
- Audit-Ready Reporting: Generates certified executive dossiers formatted for municipal comptrollers and national pension reporting bodies (e.g., PASIVOCOL / CETIL).
Mathematical Formulation
To eliminate hallucinations in financial audits, our agents delegate math to verified MCP calculation tools using formal actuarial formulas:
- Base Liquidation Income ($IBL$): $$IBL = \frac{\sum_{t=1}^{T} IBC_t \cdot \frac{IPC_{\text{target}}}{IPC_t}}{T}$$
Where $IBC_t$ represents the historic contribution base for period $t$, adjusted by the relative inflation index ratio $\frac{IPC_{\text{target}}}{IPC_t}$ over total qualified periods $T$.
- Actuarial Replacement Rate ($r$): $$r = 65.5 - 0.5 \cdot s + \Delta_{\text{weeks}}$$
Where $s$ represents the number of minimum legal wages of the worker's $IBL$, and $\Delta_{\text{weeks}}$ is the marginal percentage increase accrued for additional qualified service periods beyond statutory minimums.
How We Built It
Pensiotech is powered by a multi-agent pipeline orchestrated through Kiro:
- Model Context Protocol (MCP) Core: We built custom MCP servers that grant agents controlled access to isolated domain tools: document OCR parsing, national statutory databases, and deterministic actuarial calculators.
- Dual-Model Cognitive Layer: We leverage Claude Opus 5.5 for unstructured document synthesis, regulatory nuance detection, and edge-case legal reasoning, while using GPT-5.6 to enforce strict structural output schemas and generate audit summaries.
- Full-Stack Infrastructure: Built with a resilient Python/FastAPI backend, TypeScript/React frontend, and cloud-native document pipelines designed for high-security public sector workloads.
Challenges We Ran Into
- Document Heterogeneity: Public records from the 1980s and 1990s vary wildly in terminology, typeface, and quality. We solved this by developing an iterative agentic validation loop where agents inspect, cross-verify against known contribution years, and flag ambiguities for human confirmation.
- Zero-Tolerance for Hallucination: In fiscal auditing, an approximate number is a liability. We strictly separated LLM reasoning (interpretation of service conditions) from mathematical execution by routing all arithmetic through deterministic MCP calculation endpoints.
Accomplishments That We're Proud Of
- Reduced municipal pension audit timelines from weeks to under 3 minutes per file.
- Built an operational, full-stack platform live at pensiotech.bond.
- Successfully demonstrated that agentic workflows using MCP can tackle highly regulated, mission-critical public administration workflows without sacrificing audit transparency.
What We Learned
- How to design reliable agent architectures using the Model Context Protocol (MCP), establishing clean boundaries between generative reasoning and strict procedural computation.
- Techniques for prompt-chaining and context window management across massive administrative dossiers with state-of-the-art models like Claude Opus 5.5 and GPT-5.6.
What's Next for Pensiotech
- Direct API connectors for official territorial data warehouses (FONPET, PASIVOCOL, and CETIL registries).
- A self-service portal for municipal treasurers and public-sector legal advisors.
- Expansion of our specialized MCP tool suite to cover broader municipal debt and labor liability audits across Latin America.
Built With
- actuarial-science
- ai-agents
- amazon-web-services
- claude-opus
- document-parsing
- fastapi
- govtech
- gpt-5
- kiro
- legaltech
- mcp
- model-context-protocol
- pension-audit
- python
- react
- typescript