AI, Product & Brand
strategy for modern businesses.

UpSchool was struggling with the most common problem in any learning platform: people signed up, started a course, and quietly drifted away. Mandates weren't the answer. The real fix was making the platform itself feel responsive — like it understood where each learner was, what they wanted next, and how they preferred to be tested.

• High dropouts and low completion — users were starting courses but not finishing
• One-size-fits-all paths meant motivated learners had to bushwhack through a generic catalogue to find what they actually needed
• Assessments were rigid and MCQ-only — fine for facts, useless for evaluating reasoning, articulation, or conceptual understanding
• No social or progression layer — people had no way to see, signal, or celebrate their growth

• Built a roadmap with three reinforcing pillars: a conversational agentic learning companion, a robust dual-mode assessment system, and a badging and analytics layer — each able to stand on its own, but powerful when stacked
• Designed the agent's flow with deliberate care: the user enters topics of interest → the bot asks the user to self-select a proficiency level (Beginner / Intermediate / Advanced) → the user then answers a short set of questions to justify that level, so the system can correct for over- or under-estimation → the agent then taps the course database and generates a sequenced learning path with timelines and milestones
• Combined a strong product mindset (clear user flows, deliberate friction where it added value, removed everywhere else) with robust GenAI application (RAG over the ingested course catalogue, structured prompting for proficiency calibration, and graceful fallback paths when confidence was low)
• Specified bespoke learning paths alongside the automated ones — so admins and L&D could still curate journeys for specific cohorts where automation wasn't appropriate

• Conversational learning agent — agentic flow that takes a user's topics of interest, calibrates proficiency through a short Q&A, and generates a complete learning path with timelines and milestones grounded in the actual course catalogue
• Dual-mode assessment system — MCQ flow with detailed auto-generated report cards on completion, plus a voice-led subjective flow powered by a real-time GPT model that conducts a back-and-forth conversation with the learner and produces structured feedback notes covering strengths and areas of improvement
• Smart badging system — admin-configurable badges with custom criteria (course completion, assessment thresholds, streak-based achievements, milestone hits), tied directly to learner profiles and the leaderboard, with a curated rarity tiering so badges actually meant something
• Leaderboard & analytics dashboard — public progression view for learners; aggregate engagement, completion, and badge-velocity metrics for L&D admins to spot patterns and intervene early
• Bespoke learning path mode — admin tooling to author manually-curated tracks for cohorts where the automated agent wasn't the right fit

At a 3,000+ employee services firm, project managers were losing roughly six hours a week to Jira hygiene — drafting user stories, formatting tickets, and pulling together one-off reports for leadership that always seemed to need a slightly different cut of the data than the last one. Repetitive, high-volume work, almost always done in the same way. The kind of friction a conversational agent inside the tool itself could quietly remove.

• Junior PMs wrote inconsistent user stories and acceptance criteria — engineers asked clarifying questions, cycle time slipped
• Bespoke reporting requests from leadership ate hours: filter the backlog this way, group by that, summarise the rationale — every time slightly different
• PMs had no quick way to ask natural-language questions about their own sprint or board without manually building a JQL query

• Surveyed 40 PMs across delivery practices to map the exact toil patterns — what they did most, what they hated most, what they re-did the most
• Designed the plugin as a conversational agent — not a panel of buttons. PMs could ask in plain language, the agent would ask follow-up questions to clarify intent, and only then act
• Built the assistant to surface inside Jira's native UI — no new app, no new workflow, no training required
• Set up a lightweight evaluation harness — story quality scored against an engineer rubric, pre- and post-launch — so the team could measure adoption against actual quality

• Draft user story button on every issue — the agent generates a story from the existing description and linked tickets, and asks clarifying questions if context is thin
• Bespoke report generator — PMs describe what they need ("I need a board summary grouped by owner with risk flags"), the agent asks 1–2 follow-ups to narrow scope, then produces the report inline
• Natural-language query interface — ask things like "what's blocked this sprint?" or "which stories rolled over from last sprint and why?" and get a grounded answer

Staffing engineering teams was happening across Excel sheets, email threads, and a long tail of offline conversations between requestors, managers, and the staffing team. Project requests came in as free-form text — vague skill asks, half-specified timelines, project goals dressed up as resource requirements. The staffing team spent hours interpreting each request, chasing clarifications, and brokering between the people who needed resources and the people who had them.

• Requestors didn't know how to articulate what they needed — they spoke in project goals, not resource configurations
• Staffing decisions sprawled across tools and conversations, with no system of record and no way to learn from past matches
• Under-specified requests led to mismatches and rework — and the candidate's perspective on a fit was often lost entirely
• No feedback loop — repeated requests didn't get smarter over time, and rejection signal was never captured

• Used RAG to ground the LLM in the organisation's actual staffing history, candidate profiles, capability taxonomy, and past fulfilments — giving the model wider context with real breadth, instead of relying on general knowledge alone
• Designed a conversational, experiential flow — the agent asks the right follow-up questions to narrow focus on more apt resources, instead of demanding a perfectly structured input upfront
• Built a regeneration flow that lets the user re-enter the same conversation with refined context — each pass enhances the overall result and helps the agent zero in on the right resource pool
• Captured candidate-side signal — when a candidate isn't right for a project, they can record their reasons, and that signal feeds back into the matching model for everyone's future requests

• Conversational, agentic intake — the assistant asks targeted, contextual questions ("what's the project's primary outcome?", "is this client-facing?", "is this a known stack or are we open?") to walk the requestor toward the right resource configuration without them needing to know the taxonomy
• GenAI configuration & candidate profile engine — generates structured demand configurations (capability, sub-capability, skills, band level) and a candidate profile analysis: a per-candidate fit narrative explaining why this person matches, what their recent experience looks like, and where the closest gaps are
• Conversational regeneration — when the first match isn't quite right, the user goes through a similar (not identical) flow, with the agent narrowing further on each pass — the experiential loop is what gets to the right answer
• Two-sided feedback loop — both requestors and candidates can capture their reasoning. Requestors rate the configuration; candidates capture rejection rationale. Both signals feed the model

Marketing teams were running external communications across spreadsheets, email threads, and disconnected tools. Approvals lived in inboxes. Consent collection was manual. Posts went out without a clean audit trail of who approved what, when, and why. Compliance and brand teams had no real visibility.

• No single source of truth — same post lived in three places, each with a different version
• Email-based approvals meant decisions weren't auditable; "who said yes?" became a real question
• Consent workflows for featuring people, partners, and quotes were ad hoc — sometimes captured, sometimes forgotten
• No reliable way to enforce brand and cadence rules at scale

• Authored the full BRD covering objects, state machine, role-based access, and gate rules
• Established a single core principle: portal-only governance — every state-changing action happens inside the portal; email is only for notifications
• Designed the workflow as a controlled state machine — draft → review → consent → approval → publish → audit, with explicit gates between each stage so nothing slips through
• Modelled distinct roles with granular permissions, and specified two RBAC models — Entra group-mapped (preferred) and in-portal — with a 60-second access revocation guarantee

• Calendar planning — cadence upload with rule validation against the Content Strategy Matrix
• AI-assisted authoring — text and image generation embedded in the editor, with brand-rule preflight checks
• Governed workflow — review, consent, and approval stages with hard gates and audit logging
• Portal-native consent forms — SPOCs build, send, and finalise consent inside the platform
• Publishing engine — multi-channel content dispersion with retry logic and idempotency controls
• Audit log + retention — searchable, exportable, with legal hold support