Selected engineering work
Enterprise Agentic Analytics
Verified active years: 2026
Worked with client technology leaders and finance teams to understand financial metrics and calculation methods, then translated that knowledge into SAP-connected Python tools for AI agents. A separate fictional demo illustrates these concepts.
The problem
Finance questions depend on agreed definitions, reporting periods and source data. The engineering challenge was to turn those requirements into analytical tools that agents could use, while keeping results traceable and reconcilable.
My contribution
Client-facing AI & backend engineer
Connected client requirements with implementation: understood financial definitions and calculation methods, built SAP-connected Python analytical tools, and integrated them into agent workflows. Also contributed backend services, KPI lineage and reconciliation, and trace analysis.
Team / platform context
Contributed as part of an enterprise analytics team. This page describes my engineering responsibilities, not sole ownership of the platform. The separate fictional demo illustrates general concepts, not the employer's or a client's product.
Constraints
- Client work is confidential; this example uses fictional inputs and simplified calculations.
- Metric definitions, reporting periods and data grain must be explicit.
Illustrative skills demo. Illustrative portfolio demonstration using fictional data and simplified workflows. This is not an employer or client product. It demonstrates engineering concepts, with no live enterprise connections or AI processing. Demo outputs are not production results and do not imply employer or client endorsement.
Synthetic P6 cash tied-up duration: 42 stock holding days + 33 customer collection days − 24 supplier payment days = 51 days.
“What changed in cash tied-up duration?”
Cash tied-up duration
Synthetic P6 snapshot · matching scope and period
This example illustrates checking a calculation from fictional inputs to display; it does not document a production implementation.
From business definitions to analytical tools
My role connected conversations with client technology and finance stakeholders to the backend implementation. I worked with chief technology officers, directors, technical leads, senior developers and accountants to understand what financial metrics meant and how they were calculated.
Translate the requirement
I used those definitions to build Python analytical tools connected to SAP data, then integrated the tools into agent workflows. This gave the agent a calculation to call and a result to explain.
Make the result inspectable
My work also covered source-to-UI KPI lineage, reconciliation, testing and Langfuse trace analysis. These responsibilities connect the data source, the analytical calculation and the response shown to the user.
Evaluate answers through feedback and evidence
The team's primary operational approach is human-in-the-loop production evaluation: user feedback identifies problems to investigate, trace analysis helps locate the failing step, and source-data reconciliation and regression tests support validation of fixes.
- Capture feedback. Users submit thumbs-up or thumbs-down ratings and comments. Feedback is also recorded in Langfuse when tracing is available.
- Triage reported problems. Negative feedback is exported to a review workbook and deduplicated. The team assigns issues for investigation.
- Trace the answer. Engineers inspect the associated trace, tool calls and outputs to understand where the response diverged from expectations.
- Fix and validate. Engineers investigate and correct reported problems. Business users compare results with source-system reports or reference workbooks, while regression tests check calculations, tool behaviour and previous fixes.
Support review with automated checks
The codebase also includes an evaluator that runs questions through the actual orchestrator and scores tool-level numeric totals. Its available checks serve different purposes:
- Structural validation: checks that at least one usable numeric total was returned. This establishes presence, not correctness.
- Reference-based KPI evaluation: compares a returned KPI with an expected value within a configured tolerance.
- LLM-as-a-judge review: assesses coherence, completeness and relevance without a reference answer. This does not independently verify correctness against enterprise source data.
Historical reference-dataset runs are documented, but routine execution of every check and a continuously maintained benchmark are not confirmed. The automated evaluator scores tool totals; dedicated scoring of rendered KPI cards, charts and narrative remains future work.
User feedback guides investigation rather than measuring overall pipeline accuracy. Improvements come from engineering fixes and validation; collecting feedback is not model retraining or automatic learning. These mechanisms describe the team platform, not evaluation performed by the fictional portfolio demonstration.
Work within a wider team
The platform was team-built. My contribution was client-facing AI and backend engineering, including the translation of requirements into implementation. The page does not attribute the entire platform or its interface to me.
Engineering decisions
Keep calculations reproducible
Keep business calculations in governed analytical tools. Carry the selected definition, period and source context into the answer.
Diagnose the failing layer
Check data inputs, calculations and presentation separately when a result does not match expectations.
Validate before presenting
Reconcile tool outputs against source values and use regression checks and production traces to investigate failures.
Outcomes & evidence
- Delivered SAP-connected Python analytical tools and integrated them into agent workflows.
- Contributed backend services, KPI lineage and reconciliation, testing and trace analysis within the wider platform.
These outcomes describe my documented engineering responsibilities. The public illustration demonstrates the calculation pattern using fictional data; it is not evidence of client adoption, financial savings or production performance.
Engineering takeaway
A reliable analytical answer starts with a defined calculation. Source context, reconciliation and trace analysis make that calculation easier to inspect when the result is questioned.