Build notes / 02

Writing

I write about the systems I've built and the decisions I would make differently. Search and retrieval come up a lot.

Search and retrieval

Triloch, part 5: the search box works on older captions. I built a local path from new clips to traceable claims, then checked what it got wrong.

cctv-searchvideo-captioningvision-language-modelslocal-firsttriloch

Triloch, part 4: Qwen, Marlin and Gemma captioned the same CCTV clips. Then a bigger model and a new prompt failed to fix what they got wrong.

cctv-searchvideo-captioningvision-language-modelsqwengemmamlxlocal-firsttriloch

Triloch, part 3: AI agents built my CCTV search test tools. I ran the experiment through chat, then checked why random clips scored 68%.

cctv-searchvideo-retrievalevaluationai-agentscoding-agentslocal-firsttriloch

Triloch, part 2: motion detection asks whether something changed, and someone sitting still doesn't change. A person detector took still-time coverage to 100%.

cctv-searchmotion-detectionperson-detectionobject-detectionvideo-retrievallocal-firsttriloch

Triloch, part 1: why I'm building search over hours of recorded camera footage, running entirely on my own machines, and where it stands a few days in.

cctv-searchvideo-searchvideo-retrievallocal-firstside-projecttriloch

No embeddings and no similarity search. A policy file and a keyword scorer decide what the model sees, until two topics tie and it withholds the wrong half.

context-selectionintent-classificationllm-routingragprompt-engineeringpythonfastapi

A freelance project taught me the gap between real real-time transcription and just fast enough, and the formula that made rolling chunk transcription work.

backendsystem-designlive-transcriptionspeech-to-textdistributed-systemsfreelance

How Samvad builds a coverage report from many outlets, orders the timeline, and links its findings to source reporting.

ragllmagentspostgresside-project

I rebuilt Samvad's news retrieval around full article search and ranked evidence, then moved the model out of the retrieval path.

ragretrievalllmagentspostgresside-project

How I built Samvad's four-agent news pipeline to answer questions about Indian coverage with links back to its sources.

ragllmagentspostgresside-project

Other build notes

I checked 23 Indian Google News sitemaps against the spec: entries nine days old, a 2023 article in a two-day window, and the Go harvester built around them.

google-news-sitemapnews-sitemapsgoweb-crawlerdata-pipelinebackendopen-sourceside-project

A short reflection on Agency, Consistency, and Ego — a simple framework that’s helped me shape goals and momentum, and might do the same for you.

Self GrowthMindsetProductivity
Full size image