Love solving complex problems

github

Nivedit Jain • Peerlist

<h3>About The Token Times</h3>The Token Times is a monthly broadsheet dedicated to chronicling the significant events in the AI industry. Each edition covers the latest developments in AI models, funding, product launches, and even features an obituary page for retired models. It provides in-depth reporting on what the AI industry actually did, making complex information accessible to a broad audience.<ul><li>Monthly Publication: Receive a comprehensive overview of the AI industry's monthly activities.</li><li>In-depth Reporting: Covers models, money, launches, and retired models.</li><li>Accessible Content: Read every edition online for free or download the print PDF.</li><li>Published by Failproof AI: Brought to you by the creators of an end-to-end failure layer for AI agents.</li></ul>The Token Times aims to keep you informed about the rapidly evolving AI landscape, while Failproof AI ensures the reliability and safety of your AI agents.

the token times · a monthly newspaper on AI

Standing up an agent takes minutes. Getting real observability on it takes far longer: pick a backend, run an OTel collector, configure an exporter, wire env vars everywhere the agent runs, then figure out which spans the agent emits itself versus which you have to instrument by hand and redeploy to test. That gap is not a tooling problem. The tools are fine. The problem is that distributed tracing was designed for a different shape of program.Tracing assumes a request tree: a call comes in, fans out to known downstream services, each hop is a span, the tree closes. The structure is known before the request starts, and instrumenting means labeling edges that already exist. An agent is a loop with a model in the middle deciding what happens next, so the structure is neither known in advance nor stable across runs. Four consequences follow:<ol><li>The unit that matters is the tool call, not the network hop. When a run goes wrong, the question is what the agent decided to do, with what arguments, and what came back. Those are semantic events, and HTTP instrumentation captures none of them for free.</li><li>Identical inputs produce different traces. Two runs of the same task can take different paths and both succeed. Aggregate span statistics assume stable topology, so p95 on a span that appears three times in one run and zero times in the next is close to meaningless. You want to compare runs, not aggregate spans.</li><li>Most agent failures are not errors. A tool call fails, the agent reads the error, works around it, and finishes. Every span is green, error rate is zero, and the output is wrong. Nothing exists for a threshold alert to fire on, because the failure sits in the reasoning, upstream of anything being measured.</li><li>The payloads are the point, and they are big. Prompts, tool arguments, and tool results are what you need to read. Span attributes were not built for multi-kilobyte text, and truncating them discards the evidence.</li></ol>OTel's GenAI semantic conventions will help, but conventions standardize field names. They do not fix a request-tree model describing a decision loop, and they do not save you the afternoon spent wiring a collector.The alternative: Failproof instruments the tool call boundary instead of the network boundary. Every action becomes an event with call, arguments, result, and position in the run. Setup is npm i -g failproofai then failproofai config, plus an API key with events.add permission from the dashboard's Keys section. No collector, no exporter, no agent code changes. Existing history can be pulled in with failproofai backfill --since 7d.First hour: skip dashboards. Read three or four real runs end to end and look for repeated calls (cheap to fix, invisible until seen), swallowed errors (succeeded runs containing failed tool calls, where the interesting bugs live), and divergence between two runs of the same task, which marks judgment calls you never specified. None of these throw exceptions or page anyone.The CLI is open source; docs and dashboard at <a target="_blank" rel="noopener noreferrer nofollow ugc" class="break-all link-break text-gray-500 hover:underline" href="http://befailproof.ai">befailproof.ai</a>.

Observe Your Hermes Agent in 60 Seconds

<h3>Failproof AI is the end-to-end failure layer for teams building AI agents.</h3>It closes one loop: observe every run end to end, identify where agents fail, write a policy that stops it, and enforce it live in realtime, so the bad action never runs and the agent gets steered to what to do instead.<ul><li>Observe every run: See every prompt, tool call, and result from every agent, traced live and replayable step by step.</li><li>Identify where agents fail: Failures cluster into named modes like loops, hallucinated calls, and drift, surfacing patterns without manual digging.</li><li>Ship a policy that stops it: Use drop-in, one-line, local-only policies. Choose from 39 built-in policies or write your own. Example: on loop_detected → deny, break + summarize, restart</li><li>Enforced live, across every harness: The policy fires in realtime, preventing bad actions and steering the agent.</li></ul>Works with all your agents, including Claude Code, OpenAI Codex, Cursor, Copilot, Gemini CLI, picode, opencode, and more. A drop-in, local-only CLI with 39 built-in policies.Key Features:<ul><li>Monitoring of every prompt, tool call, and result.</li><li>Automated failure clustering into modes like loops, hallucinations, and drift.</li><li>Real-time policy enforcement to prevent undesirable actions.</li><li>Compatibility with a wide range of AI agent frameworks and tools.</li><li>Free to use with the command: npm i -g failproofai</li></ul>

Failproof AI - Build agents which never fail

Inference infrastructure that adapts to you, not the other way around. 🚀 Most inference APIs force you to make a choice: Pay a premium for instant latency, or build complex infrastructure to handle the mess of batching, errors, and validation yourself. Today, we’re launching Flexible Inference APIs by <a target="_blank" rel="noopener noreferrer" class="break-all link-break text-gray-500 hover:underline" href="https://models.exosphere.host/">ExosphereHost</a>. We built this for any volume. Whether you are running a single prompt for a quick test or processing a billion tokens for a massive workload, it is the exact same API call. Why Flexible Inference? 🛡️ 𝗚𝘂𝗮𝗿𝗮𝗻𝘁𝗲𝗲𝗱 𝗦𝘂𝗰𝗰𝗲𝘀𝘀: Stop writing retry loops. We absorb the 429s (Rate Limits) and 503s (Service Unavailable) so you don't have to. We ensure every request completes successfully, at no additional cost. ✅ 𝗕𝘂𝗶𝗹𝘁-𝗶𝗻 𝗤𝘂𝗮𝗹𝗶𝘁𝘆 𝗚𝗮𝘁𝗲𝘀: Don't pay for garbage. Configure "LLM-as-a-Judge" or auto-evals directly in your pipeline. If an output fails your criteria, we don’t just flag it, we automatically retry with feedback until it passes. You get the result you wanted, managed entirely by our system. 💸 𝗖𝗼𝘀𝘁 𝗖𝗼𝗻𝘁𝗿𝗼𝗹: Define your SLA (e.g., "finish within 10 minutes or 4 hours") and save up to 70% by trading strict immediacy for flexibility. Supports your favorite models: From open weights like 𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸-𝗥𝟭 and 𝗟𝗹𝗮𝗺𝗮 to proprietary giants like 𝗚𝗣𝗧-𝟰𝗼 and 𝗖𝗹𝗮𝘂𝗱𝗲. Start with one token. Scale to billions. The API remains the same.

Flexible Inference API | Exosphere

Exosphere: Distributed AI Workflow InfrastructureExosphere is an open-source, Kubernetes-native infrastructure platform designed to run distributed AI workflows and autonomous agents at scale. Built with Python and based on a flexible node-based architecture, Exosphere enables developers to create, deploy, and manage robust AI workflows that can handle large-scale data processing and long-running operations. 🚀 What Exosphere Can DoExosphere provides a powerful foundation for building and orchestrating AI applications with these key capabilities:<h3>Reliable AI Workflows at Scale</h3><ul><li>Infinite Parallel Agents: Run multiple AI agents simultaneously across distributed infrastructure</li><li>Dynamic State Management: Create and manage state at runtime with persistent storage</li><li>Fault Tolerance: Built-in failure handling and recovery mechanisms for production reliability</li><li>Core Concepts: Fanout, Unite, Signals, Retry Policy, Store</li></ul><h3>Smooth Developer Experience</h3><ul><li>Plug-and-Play Nodes: Create reusable, atomic workflow components that can be mixed and matched</li><li>Python-First: Native Python support with Pydantic models for type-safe inputs/outputs</li><li>Interactive Dashboard: Visual workflow management, monitoring, and debugging tools</li></ul><h3>Production Ready Infrastructure</h3><ul><li>Kubernetes Native: Deploy seamlessly on Kubernetes clusters for enterprise-grade scalability</li><li>State Persistence: Maintain workflow state across restarts and failures</li><li>API Integration: Connect to external services and APIs through configurable nodes</li></ul><h3>Built for AI Agents</h3><ul><li>Autonomous Execution: Build agents that can make decisions and execute complex workflows</li><li>Data Processing: Handle large datasets with distributed processing capabilities</li><li>Long-Running Operations: Support for workflows that run for hours, days, or indefinitely</li></ul>Whether you're building data pipelines, AI agents, or complex workflow orchestrations, Exosphere provides the infrastructure backbone to make your AI applications production-ready and scalable.🎯 Use Cases &amp; ApplicationsExosphere is perfect for a wide range of AI and automation scenarios:<ul><li>Data Processing &amp; ETL Pipelines</li><li>AI Agent Orchestration</li><li>Content Generation &amp; Analysis</li><li>API Integration &amp; Automation</li></ul>

Exosphere | AI Workflows at Scale

Discover What People Actually Want with an AI-powered agent designed to help builders, makers, and entrepreneurs find their next big thing. This innovative tool scours discussions on Hacker News to uncover real problems that people are eager to have solved. Every two hours, it identifies high-impact opportunities and shares them on social media, providing actionable insights for creators.<h3>Key Features</h3><ul><li>AI-Powered Analysis: Utilizes a large language model (LLM) to generate actionable problem statements based on current community discussions.</li><li>Real-Time Data: Leverages the Hacker News API for near real-time access to public data, ensuring up-to-date insights.</li><li>Easy Deployment: Offers a simple setup using Docker Compose, which includes all necessary services like MongoDB and Exosphere State Manager.</li><li>Customizable Notifications: Supports AWS SES for email notifications, allowing users to stay informed about new opportunities.</li></ul>Built with ❤️ by builders, for builders, and powered by <a href="https://exosphere.host" rel="nofollow">exosphere.host</a>, this project is open source under the Apache-2.0 license. Contributions are welcome to enhance its capabilities further.

WhatPeopleWant (AI Hack Finder)

Mentor: Dr Gaurav Harit, Department of Computer Science and Engineering, IIT Jodhpur. •  Worked with various versions of SRCNN to increase OCR(Google Tesseract) accuracy on scanned documents. •  Modified the loss function of various models to an OTSU-based location-specific function to better recognize text.

Simultaneous Optimisation of Image Quality an

Mentor: Dr Suman Kundu, Department of Computer Science and Engineering, IIT Jodhpur. •  Studied and worked with various software architectures and system designs (High Level and Low Level) for a few problems. •  Developed and deployed these system designs on scale for commercial purposes to solve problems like scale, reliability etc

Software Architecture to Support Enterprise A

Mentor: Dr Ravi Bhandari, Department of Computer Science and Engineering, IIT Jodhpur •  Made a democratic music control system over the local area network aimed to automate the job of party DJs. •  Including features like Voting, File Transmissions etc, supporting multiple connections at a time on RaspberryPi.

µZiK (muZiK): Democratic Party Music Control 

•  Created a RESTful Backend on Django and Django Rest Framework with PostgreSQL for Engineering Times, an online reading site for engineering students •  Developed and Planned various APIs for integration with frontend, worked on services to send mass mailing, SMS, connected CDN for efficient static files delivery.

Backend For Engineering Times (engineeringtim

Guide - Dr. Anand Mishra | IIT Jodhpur • Made a terminal-based online delivery system with all features like email verification, phone number verification, emailing, etc for the community center of IIT Jodhpur. • Awarded highest grades and selected among top 3 projects showcased on the course website. • Developed in C++ using principles of Object-Oriented Design and Analysis

CC BASKET - ONLINE DELIVERY SYSTEM

• Designed and developed backend and backend architecture based on a hybrid of microservice and event-driven for menuhub, also leading a team of 5 developers. • Using concepts like hashing, caching, messaging, task scheduling, task queue, image processing, horizontal and vertical scaling etc.

Menuhub: The Modern Menu

Guide - DR. ARTI PANDEY, IIT Ropar •  Devised an Mixed Integer Linear Programming (MILP) model to calculate best route and person assignment to a car based on distance and compatibility. •   Used IBM CPLEX Solver Python APIs to solve MILP and Google Maps Distance Matrix APIs to generate data,tested model on Chandigarh City.

CARPOOL TEAM GENERATOR

•  Designed architecture for a robust, secured, central and fast central master database management system as a freelancer. •  Designed a cloud agnostic micro-kernel architecture inspired solution and api contracts taking care of security, service-toservice authentication, latency and async tasks, process atomicity, fail-over strategy, data field level access management etc

Central Master Data Management System

Guide - Dr. Rajlaxmi Chouhan | IIT Jodhpur • Classified input music audio signal into 10 genre classes, implemented ”Musical genre classification of audio signals” by G. Tzanetakis and P. Cook. • Experimented with CNNs and KNNs to achieve an accuracy of 84%. • Implemented in Python (iPython Google Colab) using Keras and Scikit-le

MUSIC GENRE CLASSIFIER

Mentor: Dr Yashaswi Verma, Department of Computer Science and Engineering, IIT Jodhpur •  Developed a new model to transport goods to remote areas based on public participation and rewards, to save cost and fuel. •  Proved that the problem is NP-Hard and gave an Iterative Deepening Based Algorithm for it.

Load-Unload: New Approach to Transportation o

• Worked from scratch to develop components, working with customer & partners, resolving bugs and issues, completing release checklist (PLR) etc for the private preview, public preview and currently working towards general availability. • Working with technologies like: C#, dotnet, Visual Studio, Azure DevOps, PoweShell, Azure Kubernetes Services, Azure Data Explores, Apache Airflow, Astronomer.io, Azure Resource Manager, Microsoft GraphAPIs, Typescript etc.

Apache Airflow™ on Astro - An Azure Native IS

•  Developed an efficient and easy to use Python Library to interact with BunnyCDN Storage Services. •  Featured on BunnyCDN Storage APIs documentation and received $100 from BunnyCDN for helping them. • Officially listed in SDKs list for BunnyCDN • Further took this project leading and metoring a team of 2 devs (Joel and Aditya) to make a more advance and functional SDKs in DevLup Labs Winter of Code

Bunny CDN Python Storage Library

Mentor: Dr Anand Mishra, Department of Computer Science and Engineering, IIT Jodhpur •  Worked with the state-of-the-art ACGPAN and VITON for GAN and AutoEncoder-Decoder based approaches. •  Developed a pose-estimation based approach to generate sizes of various body parts.

Project Shringar: AI based Virtual TryOn

•  Contributed to Azure open source repository for Python and C# SDKs, using tools including tools like Autorest(Microsoft open source), nodejs, python, dotnet, visual studio, etc •  Developed, managed and released Python SDKs versions: 1.0.0b1, 1.0.0b2, 1.0.0b3 and 1.0.0 and dotnet SDKs versions: 1.0.0-beta.1, 1.0.0-beta.2, 1.0.0.