HackerTrans
TopNewTrendsCommentsPastAskShowJobs

ashu_trv

79 karmajoined 10 lat temu
Building @videodb_io I like building fundamental systems. Wanders in deep thoughts of technology, science, spirituality and human nature.

Submissions

Record and Replay, teach AI agents desktop workflows by showing them once

github.com
1 points·by ashu_trv·5 dni temu·0 comments

Show HN: Do Thought Streams Matter? A Benchmark of VLM Reasoning in Gemini 2.5

arxiv.org
3 points·by ashu_trv·3 miesiące temu·0 comments

Show HN: I packaged decade of video infra battle scars into tools for AI agents

7 points·by ashu_trv·4 miesiące temu·0 comments

[untitled]

1 points·by ashu_trv·5 miesięcy temu·0 comments

[untitled]

1 points·by ashu_trv·w zeszłym roku·0 comments

Live video feed for every multimodal model not just Gemini

videodb.io
7 points·by ashu_trv·w zeszłym roku·2 comments

Show HN: VideoDB – 80 % fewer hallucinations on NFL game analysis

docs.videodb.io
1 points·by ashu_trv·w zeszłym roku·0 comments

[untitled]

1 points·by ashu_trv·w zeszłym roku·0 comments

Lessons Learned Building MCP for Video Infrastructure Startup

2 points·by ashu_trv·w zeszłym roku·0 comments

Auto-Sync Your Docs, SDKs and Examples for LLMs and AI Agents

github.com
6 points·by ashu_trv·w zeszłym roku·3 comments

Ask HN: Model to Analyse Financial Transactions

1 points·by ashu_trv·w zeszłym roku·1 comments

Underwhelming MCP vs Hype

4 points·by ashu_trv·w zeszłym roku·10 comments

Benchmarking vision-language models on OCR in dynamic video environments

arxiv.org
142 points·by ashu_trv·w zeszłym roku·58 comments

Vision-Language Models vs. Traditional OCR in Video – New Benchmark

arxiv.org
6 points·by ashu_trv·w zeszłym roku·1 comments

Show HN:Video is hard: until now

github.com
4 points·by ashu_trv·2 lata temu·4 comments

Show HN: Instantly create video clips from LLM prompts

github.com
4 points·by ashu_trv·2 lata temu·5 comments

Show HN: GPT-Powered Video Retrieval and Streaming

github.com
5 points·by ashu_trv·2 lata temu·1 comments

comments

ashu_trv
·5 dni temu·discuss
[dead]
ashu_trv
·3 miesiące temu·discuss
[dead]
ashu_trv
·11 miesięcy temu·discuss
I agree. I am big fan of o3, and GPT 5 is not the same, it's like going back to GPT-3 level stupidity. It doesn't care about context, feels super dumb.
ashu_trv
·w zeszłym roku·discuss
What's really impressive here:

Model-Agnostic Infrastructure: Any AI model—open-source, proprietary, LLM, or VLM—can instantly gain real-time vision capabilities. No more waiting for Google to open their doors.

Immediate Availability: Unlike Project Astra, which is still behind Google's beta gates, VideoDB is usable today. Anyone can plug into the API, SDK, or cloud console immediately.

Openness and Developer-Friendliness: Seamless integrations with popular AI tools and frameworks like LlamaIndex, LangChain, and Hugging Face dramatically reduce the barrier to entry. Just a few lines of code and you're live.
ashu_trv
·w zeszłym roku·discuss
Yeah, we tried to solve the 1 and 3. 2nd is still an open problem. Can you share more about the MECE?
ashu_trv
·w zeszłym roku·discuss
Keeping documentation and SDK updates aligned with evolving "LLM contexts" can quickly overwhelm dev teams. At VideoDB, we've built an open-source solution—Agent Toolkit—that automates syncing your docs, SDK versions, and examples, making your dev content effortlessly consumable by Cursor, Claude AI, and other agents. Ready-to-use template available.
ashu_trv
·w zeszłym roku·discuss
A new benchmark study evaluates Vision-Language Models (Claude-3, Gemini-1.5, GPT-4o) against traditional OCR tools (EasyOCR, RapidOCR) for extracting text from videos. The findings show VLMs outperforming OCR in many cases but also highlight challenges like hallucinated text and handling occluded/stylized fonts.

The dataset (1,477 manually annotated frames) and benchmarking framework are publicly available to encourage further research.

Paper: https://arxiv.org/abs/2502.06445 Dataset & Repo: https://github.com/video-db/ocr-benchmark

Would love to hear thoughts from the community on the future of VLMs in OCR.
ashu_trv
·2 lata temu·discuss
Thanks! Added now.
ashu_trv
·2 lata temu·discuss
This open-source agent framework is like ChatGPT, but for videos. It simplifies complex video tasks like search, editing, compilation, and—best of all—generation. The results stream instantly. You can even extend the agents to suit your needs and build custom automated workflows.

The framework is fully open source and uses a VideoDB key for cloud-based video storage, processing, and streaming. It seamlessly integrates with tools like Stable Diffusion, Eleven Labs, Kling, Replicate, and more.

Looking for collaboration with GenAI audio/ video teams and feedback from amazing devs out here.
ashu_trv
·2 lata temu·discuss
It analyse the transcript, but there is no way to get back the video clip without building your own video infra. We at Videodb are solving the exact problem.
ashu_trv
·2 lata temu·discuss
LLMs are great with text, but they don't help you consume or create video clips. Checkout PromptClip - Use natural language to describe the what you want. - Instantly get video clips with the help of LLMs like OpenAI or Claude.

Few interesting prompts we tried while building it and loved the results. There's no limit to creativity with this.

*Shark Tank Videos:* [Find every moment where a deal was offered](https://console.videodb.io/player?url=https://stream.videodb...)

*Useful Gadgets* [Show me where the host discusses or reveals the pricing of the gadget](https://console.videodb.io/player?url=https://stream.videodb...)

*Huberman Podcast:* [Find details about every sponsor](https://console.videodb.io/player?url=https://stream.videodb...)

*Masterchef* [Show me the feedback from every judge](https://console.videodb.io/player?url=https://stream.videodb...)

Say goodbye to manual editing, skimming and seeking the video and hello to instant, AI-driven video consumption and creation
ashu_trv
·2 lata temu·discuss
Yeah, VideoDB is the next-gen infrastructure for videos and actually less costly than current video infrastructure.

But you can use any LLM for analysing the transcript.
ashu_trv
·2 lata temu·discuss
Build custom GPT on your video data with StreamRAG in 2 mins.This search agent find relevant moments across hundreds of hours of content and return a video clip instantly.

RAG applications are great with text, but with video they can't support simple requests like "show me where sleep improvement is discussed"