The LLM goes to fetch all the info about the particular court and then maps it out. The JEM work here all sourced from LLM. And then findings from there.

Next is Jem, judiciary entity map. Divya, Prajna. Etc. The map here how where the court is located it the systems maps, how it's connected and funded.

With its own recruitment process and promotions through out. They would already have staging areas, version control, ci cd pipe lines with all documentations.

The layered structure of leads, coordinators etcs. This totally feels like thr festival management hierarchies and timelines across and within years.

A team from IITB gets a slot here to share how they keep the projects alive when the maintainers turnover. These needs a mentoring cycle. Resobin a shared resource pool built in open for the course material.

Mohammed Shan now on Loop to Real Time from TinkerHub. Shan shows integrations of the loop on electronics. The sequential tasks add Latency. Shan proposes independent taks over bigger loop.

Lingering back here in this Hall for Jaidev Deshpande on skewed representation with skewed opinions. And then he stereotypes the online characters.

LLM judges too have biases inherent to the LLMs and they need to be tuned to the human expert. Aman proposes a full AI evaluation pipeline from evaluation to actual production improvements.

Aman makes a case for then analysing these evaluations again through ai, and get specific root cause analysis and suggestions that even domain experts can validate and suggest.

So, AI systems at scale are evaluated with LLM as judge. Separate LLM judges are preferred over a single multidimensional test. And a dashboard for the performance on all evaluators. These helps with benchmarking across models for your cases.

Aman starts with the usual phone number example but the slides are much better and improved. Making a point that irrespective of the tool call or text response, either is just next token generation.

He shares three such Agentic tools from each of these domains all from Africa and the high stakes they deal with when it comes to inaccuracies by AI.

I am fairly familiar with Calibrate, attended atleast couple of Calibrate. But now, I hope to engage with little depth and attention. Aman starts with disclaiming the claim of AI's accuracy while acknowledging their potential in health, education and agriculture.

After very heavy, very tasty and very unhealthy lunch, I am back in hall 2 for the talks. Ankesh Bharti just finishing his talk on Tiles. Next is Calibrate by Aman Dalmia.

That panel discussion appeared very popular, well attended and engaged and hence liked. But still, felt very non-answery answer. The panelist assumed that most of the attendees or addresses were directed towards the students. But most people could be early career professionals.

The once after Rukmini's invited talk was a panel discussion mostly attempting to address the anxieties around the ai enabled identity crisis.

I did not any notes in past couple of hours and going bit slow. Internet issues. But would be a good test if l let all the updates pile up here and see how a sudden surge can be handled once I am back on the network.

Rukmini in the first chapter gives the ethos of the Data for India, that is to create accountibitily loops and also to build basic enough foundation by keeping the getting into the complicated questions out of scope.

Rukhmini now for the invited 45min talk on Telling India's story through Data.

I can rush to the Hall 1 for how to build Community Projects for community from the community. Justin and Hari have finished their talks. I may have spoken already about their photo related applications.

Denny's work on the generating the grid and creating a system for design is really inspiring and feels like a lot of scope for a independent platform for non-coders and non-designers to experiment and generate visual patterns and artifact's.

Denny also refers to his north start those are very interesting pattern which are humanoids of the illustrations.

Modifier make them the grid modified in a specific way as listed above. Walker then gives you specific tracing patternned location over the grid. Drawing a function then utilises it.

Generators are simple grid layout could have range from uniform, radial grids for even any non uniform one with controlled randomness.

The problem statement that Denny is looking at is the adhering of guiding principles on the timelines. Grids had following parameters Generator, Modifiers, Walkers, Draw Function.

Denny gives a detailed and paced intro to the work at Tattle. Also, smoothly moves to the systems needed for the social media content and brand guidelines.

Denny from Tattle for P5js library for visual branding library for speedy content creation.

Ritika is talking about the UI design and Physical world offset. That includes the display, button response time, 15-30ms latency is the sweet spot for human to device interactions here.

The facilitators were collecting challenges for the community centred technology from the participants onto a board which ranged wide. But I need to rush yet again to another room.

The BoF is facilitated by Janastu on community centred technology and they have mentioned some of their initiatives like COWMesh, Community Stories and Archives, Community Radio, anthillhacks, Crafter Space and upcoming Janastu Alternative School.

Meanwhile here for genomic search engine, the existing data is enriched with additional metadata. All of this into Postgres, with custom weights for the fields for better ranking in search results.

I can run to the Community Centred Technology BoF in Room 3 for couple of minutes.

Seqout ingests data from across the dataset platforms that their own publishing/access protocols in a SQLite DB and then uses OpenAlex, ROR and pysradb. All of this runs periodically on a lab cluster.

Seqout, the platform built by the speakers, Any seach engine does data ingestion, adding metadata, indexing it and then adding extra mile features that are domain specific.

Gene expression Data and genomics are diverse and across different dataset platforms and there is no simple solution. Simple script solution doesn't scale because of the API constraints.

I shall run now for the Genomic Search engines talk in the Hall 1. The audience questions here continuing.

Another question of sustainability and revenue and bootstraping. Ayushman answers that despite the skepticism there is a way that they believe and have been getting some leads.

Another question was about to take existing Kernels like Linux core and build on top of that. And Manas kind of hints that they are sort of doing that.

The talk ends 10 mins earlier and invites to give a demo at their booth. One of the older audiences asks a question about relevant academic researches and references to them.

Manas splits that into two phases. One the is the latecy. And second is non availability of hardware drivers with thr chicken and egg problem with no users.

Manas also presents the road map for the RandD for Xevena's Kernel. Also the philosophy and capability expectations. Now Manas comes to the struggle from Hobby to Sustainable Work.

Xeneva's problems are about Scheduling, Motion to Photon (MTP) Latency. Elapsed time between the physical movement to the response from the screen. This was drop 20-40ms to 2-13ms.

Unix, BSD, Linux, QNX, Android, Chrome, VxWorks all had there own native worlds to solve the kernel related building from ground up.

After Ayushman, Manas takes over to talk about the Kernel. Mentioning that no kernel was built with XR in mind.

Latency, Comfort, Battery Heat, Fragmented Hardware and lack of standardation. But Xeneva attempts to solve the issue not building XR from ground up, than putting costumes on existing modules.

They have their own kernel for XR, which is umbrella for Virtual, Augmented and Mixed Reality.

This room is slightly ahead of the schedule. And Xenova now on surviving as FOSS. Known for their nerdy deep tech.

Priyanshu makes a case for property based testing over example based testing. But looks like the talk had started in advance and it was already wrapping up.

Now moving the UI Hall 2 for property based UI testing by Priyanshu Jain.

This was a lovely session questioninhg the security in PAN QR codes, when the obfuscation is proxy for encryption and certification while the logic is not published.

Noticed a talk in hall 3, on PAN QR codes. Scanning the QR that claims to be encrypted but it appears to be entirely local with smart obfuscation.

Hari Kiran narrated the story of the moving away from the Oracle services and the rollback plans needed when when the architecture was wrong atleast twice.

Slightly late to the Day 2. Already at the beginning of the first session: From Pull Requests to Patients. Day 2 is also opening with health tech. MeeSeva platform in Telangana running on Oracle Landscape.