Hariprasad Sudharshan, known as Hari, builds AI products for network operators and open-source tools that examine inference costs and agent behavior. A Dallas-based Senior Product Manager for Artificial Intelligence & API at 1Finity, a Fujitsu company, he advocates keeping network automation predictable and applying AI at specific points that require interpretation or tradeoffs.
From network protocols to AI products
Hari’s earlier work addressed the mechanics of keeping networks connected. With Michael Colven, he co-invented a network-protocol migration method, filed in 2015 and granted in 2017, for moving Ethernet networks from Rapid Spanning Tree Protocol to Ethernet Ring Protection Switching. The method sequences changes from the leaves of the network toward its root while keeping a protective port disabled, reducing the risk of traffic loops and disruption during the transition. A second patent, granted in 2018, covers migration from Multiple Spanning Tree Protocol to Ethernet Ring Protection Switching, with co-inventors Davin Michael Colven and Maitreya Mukhopadhyay.
His subsequent work spans telecom software, machine learning and AI product management. At 1Finity, his responsibilities extend from customer discovery through go-to-market, including decision intelligence, failure analysis and resource optimization for carrier networks.
His writing on optical network management explains why more autonomous action can make an operational problem harder to solve. An agent repeatedly adjusting optical power could delay the physical repair a degraded fiber connector actually needs. Hari argues for preserving deterministic workflows, placing AI at selected judgment points, and making its actions understandable to operators before expanding its responsibility.
Testing behavior and exposing assumptions
Hari’s open-source work makes several practical questions inspectable: what an agent actually did, what serving a model will cost, and whether an application can operate entirely offline.
Crucible: observing consequential actions. His adversarial agent-testing project examines failures involving money, secrets, permissions, identity and other consequential actions. It distinguishes a refusal an agent reports in prose from a tool action the test harness can inspect. A stricter gate rejects passes based on self-declared outcomes; when there is insufficient observable behavior to judge, a scenario can remain untested rather than receive an unsupported verdict.
Tokencalci: correcting inference-cost assumptions. Hari’s self-hosting cost planner uses workload assumptions such as request volume, token lengths, concurrency and traffic patterns. His account of its development includes a substantive correction: an earlier version overestimated fleet requirements by treating additional GPUs as though they produced proportional increases in throughput. He emphasizes utilization and staffing because downloading model weights leaves much of the work and cost of serving them unresolved.
Arivu: making offline operation a design property.Arivu is an open-source Android writing and comprehension assistant that bundles Qwen3 0.6B weights, runs them with llama.cpp and omits internet permission. Its local operation follows from how the application is packaged and what it is permitted to access.
Hari also builds smaller models to explore training and adaptation. microGPT is an educational, 1.35-million-parameter transformer trained from scratch on TinyStories to make model components understandable on consumer hardware. CricketMind adapts NVIDIA’s Nemotron Mini with LoRA for cricket-law questions and match-situation analysis. Cricket is also a pursuit away from the keyboard: Hari bowls fast in Dallas cricket leagues.
What faster building does—and does not—buy
Hari’s I Spent Two Years Building the Median reflects on discovering that other builders had independently produced calculators much like his own after two years of solo work. He calls this convergence the median: people using the same models and tools on a familiar problem can arrive at similar solutions without copying one another. Faster development still makes useful internal tools and learning projects easier to build; a functioning demo alone does not establish a lasting product advantage.
His response is a novelty budget: identify the distinctive contribution before building, put it on the roadmap and review it deliberately. For leaders evaluating demos, he proposes asking who has already solved the problem, what the team has that competitors do not, and how the application will reach users. The calculator’s later focus on enterprise self-hosting versus purchased inference illustrates the distinction he wants builders to examine: access to open weights is only a starting point, while utilization and the infrastructure between weights and generated tokens determine much of the operating cost.
Hariprasad Sudharshan’s solo-built calculator worked—and resembled calculators other people had built independently. His response was to change what he budgeted for: less attention to coding workflows, more to a specific customer problem, distribution and a difference worth building.
Separate model capability, the harness that applies it and the product customers receive. Comparing coding products does not isolate a model’s contribution.
A useful application can still be easy for others to reproduce. Internal utility, learning and commercial differentiation require different success criteria.
Before building, search for existing solutions and name the distinctive contribution. Put that contribution in the roadmap and review the effort devoted to it.
Free model weights leave inference infrastructure, operating expertise and workload utilization to account for. The revised calculator makes that enterprise decision its focus.
