Amir Haghighat is the co-founder and chief technology officer of Baseten, which builds infrastructure for running AI models in production. His specialty is production model inference: making specialized models fast, reliable, secure, and economical enough for enterprise applications.
From fraud detection to Baseten
Haghighat was head of engineering at Gumroad before working in engineering at Clover Health. At Gumroad, future Baseten co-founders Tuhin Srivastava and Phil Howes trained a credit-card-fraud classifier, then discovered that serving it reliably inside a production application was harder than developing the model itself.
In 2019, Haghighat founded Baseten with Srivastava, Howes, and Pankaj Gupta. As generative AI and openly available models proliferated, the company focused on the infrastructure needed to deploy, optimize, scale, and monitor them. Haghighat traces Baseten’s evolution toward generative AI infrastructure to that original insight: useful models create another demanding engineering problem when they encounter real users.
Specialized models create enterprise advantage. Haghighat argues that organizations differentiate themselves by combining models with proprietary data. Health insurers, for example, can use accumulated labeled records to extract diagnoses, prescriptions, and procedure codes; specialized transcription systems can better recognize clinical terminology. These applications can outperform general-purpose APIs on narrow, consequential tasks.
Latency determines whether AI feels usable. Voice applications need rapid first-token and first-sentence responses, while shared inference services often prioritize throughput. Haghighat’s co-authored analysis of GPT OSS 120B optimization details the tradeoffs among inference frameworks, hardware compatibility, tensor parallelism, cache-aware routing, and speculative decoding.
Agent economics depend on the whole workload. A single user action can trigger many model requests, overwhelming savings from falling token prices. Haghighat emphasizes compute utilization, deployment topology, and application-level unit economics when evaluating whether dedicated inference infrastructure beats per-token pricing.
Reliable inference requires an operational system. Running an open model with vLLM or SGLang does not resolve hardware failures, traffic spikes, observability, enterprise controls, or tail latency. Haghighat highlights prefix caching, fast scaling, and recovery mechanisms for mission-critical workloads, alongside private model deployments and multi-cloud capacity management.
Following Parsed’s integration into Baseten, Haghighat and his colleagues extended that vision to connect inference, post-training, and continuous improvement from production data.
Enterprise teams turn to open models when task quality, interactive latency and agent costs strain closed APIs—but owning the model also creates a production infrastructure obligation.
▶ Watch ↗