VLM Evaluation
Strong general purpose VLMs can still fail on downstream vision tasks
A small chessboard-to-FEN experiment showed Claude Opus 4.7 struggling with exact square-level localization even when it understood the board broadly.
Building the video backend for agents.
Field notes, build stories, research, and working projects from video infrastructure, retrieval, streaming, and agent tooling.
01 /
Field Notes
Field Notes
Recent production lessons
and technical notes.
Browse short notes on infrastructure, reliability, latency, retrieval, and the small fixes that matter in production.
Operations
Reuse HTTP connections in Python services
A short field note on replacing repeated bare requests calls with a shared session so Python services can reuse HTTP connections.
Frontend Performance
One CORS header that removed hundreds of milliseconds from repeated API calls
A practical note on preflight caching, custom auth headers, and why OPTIONS requests quietly dominate app latency on slower connections.
Operations
A Postgres backup is not real until you restore it
A field note from a simple offline Postgres migration: why pg_dump was enough, where read-only mode fits, and why the restore drill matters more than the dump command.
02 /
Build Notes
Build Notes
Build notes
from working VideoDB projects.
Follow the architecture choices, API decisions, and tradeoffs behind apps, demos, and agent tools built with VideoDB.
Build Note · Lalit Gupta
Bloom: Record and Search Your Screen
How I built a local-first screen recorder on top of the VideoDB SDK. It automatically indexes your audio and screen, making everything searchable by the time you paste the share link.
Build Note · VideoDB Team
Focusd: A Coach for Your Workday, Not a Time-Tracking Report Card
How we built Focusd as a local-first desktop app that records work sessions, indexes screen activity with VideoDB, and turns raw events into useful productivity coaching.
Build Note · VideoDB Team
Build CCTV Camera for Your OpenClaw Agent
How we built a VideoDB-powered OpenClaw skill that records, indexes, searches, summarizes, and clips an agent's remote desktop without changing the agent itself.
Build Note · Sankalp Nagaonkar
Deep Search: How We Built an Engine for Finding Exact Moments in Video
How we built Deep Search as a retrieval loop for finding exact moments in video using planning, indexing, validation, recovery, and follow-up state.
Build Note · Rohit Garg
Pair Programmer: live screen and audio context for coding agents
A short engineering note on turning a developer session into searchable context with VideoDB Capture, RTStreams, local events, and agent-side retrieval.
Build Note · Om
How I built Call.md: AI call intelligence on top of VideoDB
A practical build story for a local-first call intelligence app that records calls, transcribes speakers, generates live nudges, and exports structured Markdown.
03 /
Newsletter
Newsletter
Newsletters
from the VideoDB engineering team.
Read concise updates, technical notes, and build context from the team working on video infrastructure and agents.
Newsletter · May 8, 2026
The VideoDB Dispatch #001
Why agent runs need playable evidence, thinking tokens in vision models and agentic video streams.
newslettervideo-aiagents
Read the issueSubscribe
Get VideoDB Newsletter
Technical notes, build stories, and updates from our team.
04 /
Research
Research
Research at the edge
of video and agents.
Papers, evaluations, talks, and notes on video understanding, retrieval, multimodal models, and agent systems.
Research note · May 15, 2026
How to Evaluate Multimodal VLMs for Your Video Use Case
A practical workflow for evaluating video VLM setups with VideoDB and Langfuse, from task definition and dataset design to tracing, scoring, and deployment decisions.
Read the notearXiv preprint arXiv:2604.11177 · 2026/4/13
Do Thought Streams Matter? Evaluating Reasoning in Gemini Vision-Language Models for Video Scene Understanding
Benchmarks how internal reasoning traces affect video scene understanding in Gemini models, including where quality gains plateau and how tight budgets increase compression-step hallucination.
Read on arXivarXiv preprint arXiv:2502.06445 · 2025/2/10
Benchmarking Vision-Language Models on Optical Character Recognition in Dynamic Video Environments
Introduces an open-source benchmark for evaluating VLMs on OCR tasks in dynamic video environments across 1,477 manually annotated frames.
Read on arXiv05 /
Projects
Projects
Builds from the edge
of video and agents.
Apps, agent skills, hackathon submissions, and side projects.
Call.md
Turn meetings into live agent loops.
Bloom
An agentic Loom alternative where recordings become inputs for AI.
Pair Programmer
Give coding agents live screen, voice, and system-audio context.
Agentic Streams
Agents research the internet and stream clean video briefings.