AI produces numbers that are incorrect, untraceable, and unverifiable. Asking the same question twice yields different answers, and there’s no way to distinguish right from wrong outputs. This unreliability poses serious risks across industries like healthcare, legal, insurance, government, and finance where numerical accuracy directly impacts outcomes.

Through conversations with 137 financial firms - including private equity, hedge funds, and investment banks - a consistent pattern emerged: everyone wants to use AI, but nobody trusts it. As one managing director stated, “I can’t put a number in front of a client if I can’t show where it came from.”

The problem is fundamental: AI hallucination can’t be solved by making AI smarter alone. Training data improvements and guardrails reduce errors but don’t eliminate them.

Kepler’s solution is distinct: prevent AI from producing numbers altogether. When users ask Kepler questions, AI interprets intent while deterministic code retrieves data and performs calculations. AI and code operate in separate lanes, eliminating the possibility of AI-generated numerical errors. Every figure traces to its source; every answer is reproducible.

The team draws expertise from Palantir, where I spent seven years building data infrastructure, and Citadel, where co-founder John worked developing user-facing data products. Team members come from Meta, Bloomberg, and Stanford, backed by founders of MotherDuck, dbt, and figures from Facebook AI Research and OpenAI.

The company invokes Johannes Kepler’s historical use of trusted astronomical data to make groundbreaking discoveries, positioning trustworthy data infrastructure as foundational for AI advancement.

Throughout my career, I have dedicated myself to creating tools, products, and technologies that help people effectively utilize their data. My passion lies in developing products that enable users to efficiently and scalably gain maximum value from their data.

My journey in understanding the intricacies of data and its potential began at Palantir Technologies, where I began working on search and indexing products. As data volumes grew, I focused my efforts on solving some of Palantir customers’ core problems across the financial and defense verticals before leading customer focused compute teams. After Palantir, I served as CTO at Veraset, a cloud-based data-as-a-service company. Veraset delivered high-quality, large scale data to a number of enterprises and grew to 15M ARR before being acquired. Following this, I joined Citadel Investment Group as the Head of Business Engineering of Ashler. In that role, I managed crucial data operations, including overseeing data pipelines, investment platforms, data lakes, and the software and data engineering teams responsible for them.

Currently, I’m working on building a new company in the data platform + AI space.

What are you building?

I am developing a prototype to showcase how Large Language Models (LLMs) can revolutionize a major public institution in New York City. This institution’s data is currently scattered across various systems (not all of which are organized). The solution I’m creating will integrate external APIs, fundamental coding elements, user-generated commands, and the intuitive interaction with a large language model. This will allow individuals to pose complex inquiries in everyday language. My objective is straightforward: by highlighting the recent progress and sophisticated tools available in our field, I aim to initiate this enterprise’s digital transformation.

What technology have you been using so far?

Before adopting CALM, I utilized a blend of LangChain/LlamaIndex for “routing,” Chroma for my Vector Database, and OpenAI’s models for my cloud-based LLM tasks. For on-prem LLMs, I used Databricks Dolly/MPT models. Each of these technologies excelled in their specific functions, but there was a notable lack of cohesive integration among them. As such, my experience largely involved playing a guessing game with temperatures and prompts to steer a conversation into the fixed pattern that I wanted. Additionally, the unpredictable nature of the models meant that maintaining a natural, fluid conversation often had to be sacrificed in favor of rigid rules. This constraint resulted in a user experience that was less enchanting, spontaneous, and as early customers described…less magical.

Describe CALM in 3 words:

“Transformative, Extensible, Simple.”

What are your biggest challenges in building and maintaining an AI assistant?

The majority of tools in the industry are developed with a primary emphasis on the tool itself, rather than creating a holistic user experience. This focus can make the development process challenging and unwieldy for developers who have to navigate and integrate various backend elements, models, and systems in an effort to achieve a smooth user experience.

Furthermore, there is a prevalent ‘all-or-nothing’ mentality in the utilization of Large Language Models (LLMs). This approach suggests that LLMs should either be responsible for the entire workflow, from data analysis and code generation to producing readable output, or not be used at all. Finding products that leverage the strengths of LLMs while also providing developers with flexibility and choice, especially in scenarios where LLMs may not be the most suitable option, remains a challenge.

Finally, AI assistants are fundamentally task oriented. These tasks generally have multiple steps and workflows associated with them. Actually visualizing these steps and ensuring they work under a variety of conditions is a challenging part of the AI assistant development process.

Which functionality of CALM helps the most to resolve the challenges of building AI assistants?

I don’t think I have ever seen a product that has increased my developer velocity more than Rasa Pro with CALM. CALM had 2 features that, in particular, stood out to me.

Declarative YAML flows - CALM’s declarative user based flows puts developers right in the seat of the user. By describing the exact user flow, coupled with validation and data gathered, developers not only can clearly and deliberately articulate the ideal conversation flow, but they can also visualize and extend it. It allowed me to start my development work with an articulation of the happy paths, so I was able to build quickly and with guardrails.

Conversation Repair - In the ideal world, users would follow our predefined conversation flows to the letter. However, real-world interactions are rarely so linear. Users often diverge from the set path, whether to ask questions, explore tangents, or jump into different flows entirely. This is where the magic of conversation repair comes into play. It’s a foundational element that gracefully handles these deviations, allowing users the freedom to stray from the main path without getting lost. They can ask something off-topic, follow a different thread, and then seamlessly return to the original flow. I’ve found that integrating conversation repair into chatbots is crucial for mimicking human interaction. It’s not just about guiding users through a set journey; it’s about creating an experience that acknowledges and adapts to their natural conversational behavior. This flexibility is key to making chatbots more relatable and engaging, transforming them from mere tools into conversational partners.

What prompted you to explore CALM?

I quickly realized the power of a structured approach to chatbot creation while playing with Rasa’s open source tool. This platform enabled me to set specific rules for extracting entities and establish policies for guiding the flow of conversations. Working with Rasa’s open source solution was a game-changer for me as a developer. However, during this process, I began to recognize a limitation: while the Natural Language Understanding (NLU) models were powerful, they didn’t quite capture the more dynamic, almost magical interaction quality offered by Large Language Models (LLMs). This realization led me to explore how LLMs could be integrated with Rasa. This is when I discovered CALM, which combines the power of LLMs while maintaining the controls of an NLU based-approach.

The Rasa Pro Developer Edition from Rasa gave me the chance to explore the full scope of the product functionality and build a prototype.

Link

https://rasa.com/blog/navigating-calm-s-benefits-technical-insights-from-vinoo-ganesh/

My name is Vinoo and I’ll be your guide, writer, and (likely) ranter throughout this series.

0. What this is

Recently, I had a friend recommend Austin Kleon’s Show Your Work. I read it over the course of a short plane ride and realized… it’s time to start showing my work. Publicly.

This blog series, at its core, is a collection of things learned while building analytical tools, creating data products, and consuming data products. Mostly, it’s intended as a technical series for technical audiences, but the direction remains uncertain.

My goal is to extract meaningful and applicable lessons from the challenges I’ve faced working in the data ecosystem and democratize the best practices I’ve learned along the way.

1. Why this, why now

The data industry experienced a golden age. Cloud migration accelerated. Organizations became data-driven. The modern data stack seemed powerful and limitless. Capital flowed freely, making expensive data platform investments seem justified.

Economic conditions shifted dramatically. Layoffs multiplied. Companies cut visible expenses — office space, travel, contractor fees, and staff. However, their compute bills are increasing by an order of magnitude year over year.

The compounding cost of clouds is an existential threat for enterprises without capital assessment mechanisms. Data practitioners who understand cost optimization methodology could literally save their companies from unexpected bills — or worse.

2. Who may benefit from this

Y Combinator recently distributed guidance to founders emphasizing cost management. This series targets anyone seeking to optimize data and technology spending during constrained periods.

3. Why me

I’ve spent over a decade working across tech and finance, building analytical tools, brokering data, and optimizing internal tech stacks. I’ve made numerous mistakes along the way, and I’ve learned from them.

4. Where I’ll start

The series will cover data storage, management, analysis, visualization, case studies, and best practices based on personal experience and industry examples.

I’m thrilled to have you join me on this journey, and I hope you’ll find value in the content we’ll be sharing.

Code

https://github.com/apache/parquet-site/commit/3563721676b364b767058a953f2bcc3e2c0c4b09

Link

https://parquet.apache.org

An engineering undergraduate student at Washington University in St. Louis helped create and launch a mobile app that helps students track the campus circulator shuttle. Vinoo Ganesh, a senior majoring in computer science in the School of Engineering & Applied Science, developed the real-time tracking app, titled “WUSTL Circulator.” The app also shows the circulator’s route and stopping points along with a full schedule that students can browse.

Ganesh developed the application in collaboration with staff from WUSTL’s Parking and Transportation Services and Information Services & Technology.

Ganesh says he took a class on building apps for iPhones and was looking for a way to translate that knowledge into a real-world project. Several friends on campus suggested that knowing when the circulator shuttle will arrive would be a big help.

“The reason I did this is a service to the Wash U students,” he says. “All the administrators got completely behind the effort.”

John Bailey, assistant director of Student Technology Services in IS&T, coordinated the effort and worked closely with Ganesh on the project.

“Vinoo’s idea and prototype app were fantastic,” Bailey says, adding that some additional work was necessary.

The university needed to purchase and install GPS units on the shuttles, set up a server application to feed location data to the app and extensively test it, he says.

Parking and Transportation provided funding for the project, which allowed the university to purchase the necessary GPS units and associated data plans. IS&T provided the server infrastructure to drive the app.

Oliver Jones, a graduate engineering student and a student staff member in Student Technology Services, was instrumental in developing the server mechanisms that collect GPS data and feed it to the app.

As with other student-developed university mobile apps, Ganesh partnered with IS&T during the publication process. IS&T has established guidelines and procedures for reviewing and publishing student-developed apps and already has helped publish 11 apps that were developed in part, or in whole, by students, Bailey says.

The circulator app for iPhones, iPads and iPod touch devices went live on Jan. 23 and is available for free through the iOS app store. An Android version also is in the works and could be ready in coming weeks. Jones is working on that project as well.

As an indication of the app’s demand and popularity, Apple’s stat tracking tool indicated the WUSTL Circulator app had 1,102 downloads on its first day of availability – three to four times the typical number for the first day of a new WUSTL app, Bailey says. By now, it has been downloaded more than 2,300 times, he says.

Ganesh says it’s great to add publishing an app to his résumé, but he also hopes to encourage other student developers to pursue their ideas, saying the university’s support was instrumental in helping his project succeed.

Parking and Transportation hopes the WUSTL Circulator app encourages more people to use the shuttle and improves the experience for regular student riders.

Link

https://source.wustl.edu/2013/02/new-mobile-app-helps-students-track-campus-shuttle/