Selected tools and process improvements.

Understanding how each venture in the portfolio was progressing meant manually reading through journey documents, meeting transcripts, and session feedback. It was time-consuming, and limited in how much of the portfolio it could realistically cover.

Give the program team a single, reliable view of venture health — one that scores every venture the same way, draws on all the available evidence, and can grow as the program wants to track new signals.

• Built a Streamlit dashboard that reads live from SharePoint through the Microsoft Graph API, pulling in journey documents, transcripts, session records, and feedback.
• Used an LLM to extract signals from those documents and tag them Green, Amber, or Red, then scored each venture on momentum (founder engagement and sprint progress) and investment (resources committed), with the overall RAG set by the weaker of the two.
• Made cross-referencing a rule: every venture is scored using its notes, its own folder files, and the common documents together — never a single source, and never skipped because one source was blank.
• Calculated NPS-style signal scores by formula at render time instead of calling the model for each venture, and structured the code to read all documents once, saving processing time.

Portfolio health moved from document-by-document reading to one dashboard: RAG status by venture and hub, venture- and portfolio-level NPS, mentor session insights, and a synthesis of value delivered. The program can now add new signals to its health tracking, which wasn't practical with the manual approach.

Matching mentors to ventures across a national, multi-hub network relied on ad hoc, manual selection — hard to repeat and hard to audit.

Create a repeatable, auditable way to shortlist the right mentor for a venture's specific need, with clear rules about who should and shouldn't be recommended.

• Built a Mentor Search AI tool using semantic search over mentor bios, expertise, industry, and operator experience.
• Designed a two-tier classification — Tier 1 where industry match and hands-on operator experience are both confirmed, Tier 2 for partial matches — along with a 10-point scoring rubric.
• Wrote standing rules into the process, such as one mentor owning each session and exclusion criteria for conflicting profiles.
• Standardized legacy expertise and industry labels into a consistent taxonomy so matching works from clean data.
• Tested the tool's suggestions against independent AI assessments to surface systematic issues to fix.

Mentor selection became a repeatable process: every shortlist carries a tier and a score that someone else can review and challenge, and the standing rules apply the same way each time. The standardized taxonomy carries across the mentor platform, not just this tool.

A GPT-based agent was generating outputs from source transcripts that were meant to reach mentors and ventures. Before they did, the open question was how faithfully those outputs reflected what had actually been said.

Check the agent's outputs against the source transcripts and find out whether, and how, they departed from them.

• Ran a structured, manual gap analysis comparing the agent's outputs against the source transcripts.
• Looked for patterns rather than one-off errors, and identified systematic hallucinations such as invented deadlines and fabricated deliverables.

The issues were caught before those outputs reached mentors or ventures — including deadlines and deliverables that were never in the source.

Invoices from experts and service providers had to be reviewed, validated against their contracts, and passed to Finance for processing. The earlier approach looked only at the total amount and the contract validity dates before approving — which allowed duplicate billings to be processed.

Review and validate each invoice and send it to Finance for processing, with every invoice mapped to its contract — and close the gap that had let duplicate billing through.

• Designed a process that logs every call and maps it to the invoice and the contract it belongs to.
• Reviewed each invoice's format and confirmed that all related supporting documents were in place.
• Validated each invoice against three checks: (1) every call listed has actually happened and has not been billed on any other invoice; (2) the invoice total does not exceed the contract amount; (3) the total capacity billed does not exceed the contract capacity.
• Sent validated invoices to the HoD for approval, and then to Finance for payment processing.

Approval no longer rests on the invoice total and validity dates alone. Each invoice is now checked call by call against a log, and against the contract’s amount and capacity limits, so duplicate billing is tested for before an invoice reaches the HoD and Finance.

Four National Heads handled venture sourcing, and each kept their own tracker. Counts could be adjusted without any independent check, so management did not get a single number when reporting.

As PMO Head, give management one reliable sourcing number — through a process change the National Heads would accept.

• Designed a simple process change: one sourcing tracker, maintained at a central location, in place of four separate ones.
• Put the PMO team in control of the central tracker, with check points to make sure the data is maintained properly.
• Convinced all four National Heads to adopt the change.

The sourcing tracker is now maintained centrally, with the PMO team holding control and check points. Management reports from a single set of data that is maintained properly, instead of four separate trackers.