Prism
The Story

One engineer. AI tools.
A production platform.

Prism is a video intelligence platform that processes real content at scale. It was built by one person using AI as a force multiplier. This is how — and what it means for your team.

130K+
Scenes Analyzed
~5 min
Per Hour of Video
57%
Cost Reduction (v4 to v5)
1
Engineer

A problem that should have taken a team of 12

A major media network needed to unlock the value trapped inside thousands of hours of video content. They wanted semantic search across their entire library, automatic actor identification across videos, intelligent ad break detection, and cross-platform analytics linking content attributes to audience behavior on YouTube, Facebook, and Instagram.

The conventional approach: hire a video engineering lead, 2-3 ML engineers, a data engineer, a backend developer, a frontend developer, an infra engineer, and a product manager. Conservative estimate: 12 people, $3M+ per year, 12-18 months to production.


AI as a force multiplier, not a toy

I didn't use AI to generate boilerplate. I used it as a thinking partner across every phase of the work — architecture design, algorithm selection, trade-off analysis, implementation, testing, operations, and documentation. The result was building at a pace and quality level that would normally require a full team.

Architecture & Design

AI explored trade-offs (pgvector vs. Pinecone, RRF vs. weighted scoring) in hours, not weeks

Implementation

6+ Cloud Run services, a Next.js frontend, dozens of database functions — all AI-assisted

Operations

Self-healing patterns: stall recovery, circuit breakers, capacity regulation — no dedicated SRE

Cost Optimization

AI analyzed processing telemetry and identified architecture changes that cut costs by 57%

The key insight: AI doesn't replace engineering judgment. It amplifies it. You still need to know what to build and why. But the speed at which you can explore, validate, implement, and iterate becomes radically different.


Not a prototype. A production system.

A two-phase map-reduce processing engine that ingests raw video, detects shot boundaries using triple intersection (visual cuts + audio silence + black frames), chunks intelligently at narrative boundaries, then processes in parallel across hundreds of serverless workers. A single multimodal AI call per chunk extracts scenes, people, objects, text, speech, themes, mood, tension, and narrative structure — simultaneously.

A global unification layer that resolves entities across an entire video using a million-token context window. "The expert" in scene 3 and "Dr. Smith" in scene 47 become one person. Transcripts merge from multiple sources. Faces detected during processing can be enrolled and labeled retroactively — identify someone once, and they're recognized across the entire library without reprocessing.

A six-phase hybrid search pipeline — LLM query decomposition, faceted filtering, BM25 keyword + vector similarity search, Reciprocal Rank Fusion, local cross-encoder reranking (zero API cost), and video-level scoring. Compositional queries like "makeup scene in a car" actually work because both concepts must co-occur. Every result includes full pipeline explainability.

Cross-platform analytics dashboards connecting scene-level content intelligence with YouTube retention curves, Facebook engagement, Instagram reach, and ad revenue data. Not separate dashboards — fused intelligence that answers questions no single system can.


Then I built the tool that makes the team faster

Building a platform is one thing. Making an engineering team more productive every day is another. So I built Jeeves — an AI-powered Slack and Teams assistant that acts as institutional memory for an entire engineering organization.

Ask Jeeves to debug an issue, and it searches past Slack threads, JIRA tickets, GitHub PRs, and Confluence docs — then tells you whether anyone has solved this before and how they fixed it. Ask it to create a ticket, and it reads the entire discussion thread, extracts the problem statement and acceptance criteria, and shows a preview before creating anything. It learns from corrections: cancel a ticket preview and change the project, and Jeeves remembers your preference next time.

7 Core Actions

JIRA creation, thread summary, debug lookup, multi-source search, video QoE analytics, channel digest, conversational help

6 Integrated Systems

Slack, Teams, JIRA, GitHub, Confluence, MediaMelon — unified search across all of them

Learning Memory

Tracks user preferences, detects corrections, builds channel patterns — improves accuracy without retraining

Safety by Design

Five-layer guardrails: read + create only. Impossible to delete or modify existing resources. Preview before every action.

The debug action finds past resolutions in 5-6 seconds using just one LLM call — heuristic keyword extraction handles the rest. A question that used to take 2-3 hours of cross-team investigation now takes seconds.


Your team can do this too

Prism proves that one person with AI tools can build what used to require a team of 12. Jeeves proves that AI tools can make an existing team dramatically faster — eliminating repeated investigations, automating routine work, and surfacing institutional knowledge that would otherwise be lost in chat history.

The same methodology can transform how your engineering team operates. A 5-person team using AI-augmented workflows can produce the output — and the quality — of a team three times its size.

I bring 18 years of engineering leadership in media technology — from ad platform architecture serving a billion video views annually to AI/ML systems built from scratch. I work with early-stage teams to embed these workflows into how they build, review, deploy, and operate software. The engagement is typically 90 days: audit your current workflow, implement AI-augmented processes tailored to your stack and team, and measure the before-and-after impact so you have hard numbers, not promises.

Make your team
unreasonably effective

If your team is talented but stretched thin, and you're wondering how to do more without hiring faster, let's talk.