Work that shipped.

People today sit through back-to-back meetings across completely different contexts — an internal team meeting on Teams, then a client call on Google Meet, then a sales call on Zoom. There's no single place that shows what happened across all of them, regardless of provider. Context gets lost between calls, and there's no fast way to prep for a new meeting using history from a similar past one with the same people or topic.

Solo users with a high meeting load — consultants, freelancers — and teams like sales and marketing who are in constant back-and-forth conversations where details are most likely to get missed.

Rather than depend purely on a single third-party transcription API, built an in-house transcription and summarization engine so one centralized system can handle meetings from any provider — using the provider's native transcript when available, and falling back to Visibli's own meeting bot transcript when it isn't.

Unlike single-channel meeting-note tools (e.g. Fireflies, Otter) that only email a summary, Visibli gives one unified interface across every meeting provider, lets users click back into any previous meeting's context, and pushes the full output directly into the team's communication channels — not just an inbox.

Sales and CS reps were losing 30–45 minutes after every customer call: writing summaries, extracting action items, updating CRM fields manually, drafting follow-up emails. The data was inconsistent, often missed, and never reached the right place.

Leadership had zero visibility into what was actually being said in customer conversations — the most valuable data in any B2B company.

Owned end-to-end product definition: user research with Sales Execs and CS managers, persona creation, journey mapping, PRD writing, and engineering handoff across a 10-member team.

Defined the full product specification: bot behaviour, transcription pipeline, LLM prompt architecture for structured insight extraction, CRM field mapping, and follow-up email generation logic.

The company ran four CRMs in parallel across different teams. One customer could exist as four different records with conflicting data: different names, deal stages, health scores. Nobody could confidently answer "who is our best customer?"

Reporting was manual and monthly. Leadership decisions were based on whichever CRM the presenter happened to use that day.

Contributed to the design of a Canonical Data Structure Layer (CDSL) that normalised customer lifecycle data from all four CRMs into a unified platform model. Defined entity relationships, field mapping standards, deduplication rules, and conflict resolution logic.

Managed CRM structuring, data mapping, and workflow configuration across all four platforms. Coordinated phased rollout with engineering and per-team stakeholder alignment.

Revenue teams were context-switching constantly — pulling data from 6+ tools to answer questions that should take seconds: "What's pipeline health this week?", "Which deals are at risk?", "What happened on the last 3 customer calls?"

A single monolithic AI assistant wasn't the answer — it would be too slow, too generic, and impossible to test. The solution was a specialised agent network, each expert in one domain.

Designed the agent taxonomy: identified 25+ distinct contexts where AI replaces manual lookup. Wrote product specs for each agent — defining inputs, outputs, data sources, edge cases, and inter-agent handoff contracts.

Led sprint planning and dependency management across the 10-member engineering team to ship agents iteratively without blocking each other.

Enterprise SaaS platforms don't share data cleanly. Each has its own API schema, authentication model, rate limits, and webhook behaviour. Building a product that unifies them requires carefully designed integration specs — not just API calls.

I owned the integration requirements layer: defining what data flows where, in what format, with what fallback behaviour, and how conflicts between systems get resolved.

Integration workflow and requirements documents for each platform. Auth flow designs (OAuth, JWT, SSO, RBAC). Data field mapping specs. Error handling and retry logic definitions. Coordinated with engineering across Dev → Staging → Production environments via Azure DevOps.

Also contributed to SOC 2 and ISO 27001 readiness: audit trail requirements, access control planning, and security workflow reviews for compliant integration architecture.

As the AI agent ecosystem grew past 25 specialized agents, there was no single place to see which agents existed, whether they were healthy, what they were allowed to do, or who had approved their actions. Agent sprawl created a governance blind spot: leadership couldn't answer "how many agents do we have, how well are they performing, and can we prove every action they took was compliant?"

Designed the full fleet-management console: the Agent Catalogue (registry of every agent with health, trust score, and version), a Configuration layer for per-agent parameter control, an Operations view surfacing cross-agent insights ranked by severity, an Automation layer for bulk scheduling across the whole fleet, a Performance dashboard for tenant-wide system health, and the Evidence and Governance consoles that make every agent's run history, policy compliance, and approvals fully auditable.

Also defined the per-agent detail structure — Overview, Dependencies, History & Runs, Performance, Configuration, Scheduling, and Governance — so this same seven-part lifecycle applies identically to all 26 agents.

Pipeline data lived across three disconnected concerns: account structure and territory governance, forecast accuracy and quota attainment, and the actual day-to-day views reps and managers use to work deals. Without a unified design, each of these risked becoming a separate tool with inconsistent data and no shared governance model.

Designed and shipped three connected sub-systems under one pipeline module. Accounts & Hierarchy: an interactive account tree with node-level relationship views, governed change-request workflows for hierarchy edits, and roll-up intelligence with currency and double-count attribution handling.

Forecast Variance & Targets: a full forecasting workspace — quota-per-rep tracking, variance decomposition by driver, governed manager adjustments with mandatory evidence, and deal-level drilldown with full audit snapshots.

Boards, Lists & Calendars: the operational workspace itself — a governed Kanban board with WIP limits, a synced list and calendar view, a shared filter/segment builder, and cross-team collaboration with commenting and shared views.

Forecast planning, scenario modeling, and variance explanation typically live in disconnected spreadsheets with no shared governance, no audit trail, and no way to trace a number in the final forecast back to the deal or driver that produced it.

Designed an eight-tab governed planning workspace: a live Planning Grid with formula controls and configurable time granularity; Versions & Compare for full version lifecycle and baseline diffing; Scenario & Simulation with branching, a promotion pipeline, and Monte Carlo-style probabilistic modeling; Drivers & Adjustments for bulk changes and quota overlays; Collaboration & Review for in-context commenting and conflict-safe concurrent editing; Governance & Access with RBAC, lock/freeze windows, approval gates, and data masking; Evidence & Lineage with immutable, cryptographically hashed audit trails; and Operations & Interop for import/export, performance, and error handling.

Different revenue stakeholders need fundamentally different views of the same underlying data — a CRO cares about pipeline coverage and forecast risk, a CFO cares about margin and cash conversion, a RevOps lead cares about deals needing attention. Building three separate dashboards would fragment the data model and triple the maintenance burden.

Designed a single Command Center with seven tabs — Command Center, Persona Lens, Revenue Signals, Adaptive Widgets, Governance & Visibility, Runtime & Reliability, and Evidence & Audit. The signature feature, Persona Lens, reshapes the same underlying widgets and KPIs based on who's viewing — CRO, CFO, or RevOps — with full audit logging of every persona switch.

Underneath that sits a governed widget system (Adaptive Widgets) letting any user personalize their own layout from a governed catalog, and a full RBAC and row-level-security layer controlling exactly what each role and hierarchy level can see.

Managers were spending hours pulling data from multiple systems to prepare for 1:1s and identify coaching priorities, with no consistent way to know which rep most urgently needed attention or what specifically to coach them on.

Designed a single AI-driven coaching workspace built around a clear reasoning chain: Who Needs Attention (AI-prioritized by impact and urgency), What I Learned (patterns and key observations), What I Recommend (a best-option plan with simulated alternatives), and What I Prepared For You (auto-generated coaching plans, meeting guides, and goal proposals).

Underneath that sits an Agent Command Center showing model confidence and a simulation preview of executing the plan versus doing nothing, an Agent Memory layer that tracks coaching themes learned over time, and a Multi-Agent Collaboration view showing which connected intelligence agents contributed to each recommendation. The whole system is wrapped in a Governance Command Center requiring manager approval, policy compliance, and full audit logging on every action.

Jalsa Events ran multiple concurrent events with no centralised visibility system. Vendor coordination happened over WhatsApp, CRM updates were manual, and execution status lived in the event manager's head.

This was where I first built systems for managing people, timelines, and information across multiple concurrent workstreams — the same skills that later powered AI product execution.

Centralised execution tracker for concurrent events. Operational reporting system giving leadership real-time visibility. Vendor communication frameworks. CRM update workflows. Cross-team coordination processes.

The systems reduced time-to-decision on vendor issues and gave leadership a single view of event health across all active projects.