Executive Summary
From "where is that file?" to instant retrieval
The client's marketing, sales, and enablement teams depended on a small number of people who happened to know where things lived in a sprawling, inconsistently named media library. I led the architecture, build, and governance program for a voice-first media intelligence platform that turned that tribal knowledge into a governed, searchable capability — reducing retrieval time from minutes to seconds and freeing the business from single-person dependency.
The Business Problem
The client's critical marketing and sales assets lived in deep, inconsistently named folder trees across shared drives. Finding the right image, video, or document depended almost entirely on knowing who to ask — usually one or two long-tenured team members — rather than any searchable system. This created adoption drag across marketing, sales, and enablement, and a real risk of lost productivity whenever those key people were unavailable.
A baseline audit at engagement kickoff quantified the exposure: average asset retrieval time of 6.2 minutes when successful, and a meaningful share of requests escalated directly to two "librarian" team members who together fielded an estimated 40 escalations per week. Leadership had no visibility into how often assets were being recreated from scratch simply because an existing one couldn't be found — a direct, if invisible, cost to marketing efficiency.
The Solution
I designed and delivered a media intelligence platform with governed indexing, voice-first queries, and ranked retrieval. A scheduled crawler keeps the index current across every folder, a speech recognition layer tuned for the client's domain vocabulary understands natural phrasing, and a ranking engine combining fuzzy matching and similarity scoring returns previewable results in seconds rather than minutes.
The platform was built to keep the index verifiably current, not just fast: every crawl run is logged, and low-confidence matches are surfaced distinctly from high-confidence ones so users can trust what they're seeing. That reliability signal — knowing whether a result is a strong match or a guess — was what got skeptical long-tenured staff to actually adopt the tool instead of falling back on asking a colleague.
How It Works
Voice
Speech recognition tuned for domain vocabulary and natural phrasing, not generic dictation.
Indexing
Scheduled crawlers extract metadata and keep the index fresh across every folder automatically.
Ranking
Fuzzy matching and similarity scoring tuned for speed and confidence, not just raw keyword hits.
Architecture & Governance
The diagram below traces how a spoken query becomes a ranked, previewable result, with an automated confidence check distinguishing strong matches from low-confidence guesses before anything reaches the user.
System Architecture — Query-to-Retrieval Flow
Five stages · one automated confidence checkpoint
Natural spoken or typed request for an image, video, or document
Domain-tuned speech recognition maps natural phrasing to searchable intent
Scheduled crawlers extract and refresh metadata across every folder automatically
Fuzzy matching and similarity scoring distinguish strong matches from low-confidence guesses
Previewable image and video results delivered in a fast, responsive grid
Index freshness log runs beneath every stage — every crawl run is timestamped so the team always knows how current the index is.
- Speech recognition layer for voice input, tuned with domain-specific vocabulary rather than generic dictation models.
- Indexing pipeline that scans folders, extracts metadata, and updates the index on a fixed schedule.
- Search and ranking engine combining fuzzy string matching and similarity scoring for speed and confidence.
- Responsive web UI with a preview grid for images and videos, so users confirm a match before opening it.
Business Impact & ROI
- Cut average asset retrieval time from 6.2 minutes to under 25 seconds by replacing folder navigation with governed, ranked search.
- Reduced escalations to the two key "librarian" staff members by 76%, removing a single-point-of-failure dependency on tribal knowledge.
- Lowered duplicate-asset creation by giving every team a reliable way to confirm whether an asset already existed before recreating it.
- Delivered a fully costed ROI model presented to the executive team: total program investment recovered in approximately 7 weeks of production use, driven by time savings and reduced duplicate work, with a 4.4x return in year one.
- Improved adoption by making discovery frictionless enough that even infrequent users could find what they needed without help.
- Gave leadership visibility into search patterns and gaps in the media library for the first time, informing future content investment.
My Role
I owned this engagement end-to-end as the client's AI/ML product leader, operating at CTO scope: I built the business case and ROI model presented to the executive team, defined the operating model for indexing and retrieval, and set the confidence-signaling standard that distinguished trustworthy matches from guesses — the detail that ultimately drove adoption.
I also aligned the UX to how marketing, sales, and enablement teams actually searched for assets, rather than forcing them to learn a new mental model, and shipped the voice and ranking workflows end-to-end. The result was a platform the client's leadership could scale across departments as a permanent knowledge asset, not a one-off search tool.