Maitai is a $50 to $200/mo control plane that indexes production LLM traffic and auto-corrects bad output in real time with Sentinels.
Founded in 2024 by Christian DalSanto and Ian Hoegen, Maitai replaces ad hoc prompt patching with a managed control plane: a Portal and SDK that trace every production call, catch bad output live, and turn those traces into fine-tuned, distilled models over time. Built for regulated, latency-sensitive, or irreversible-action workloads like voice agents and customs classification.
Sentinels that intercept and correct bad output inline, then feed the correction back into the next fine-tuning cycle automatically.
The Sentinel-plus-fine-tuning loop is a real product wedge, but the moat is thin: incumbents like LangSmith, Arize, and Braintrust have far more funding, integrations, and reference customers, and a customer's own trace data is the only true lock-in Maitai holds.
Main trade-off
No public third-party reviews and a very small team, roughly six people, to support enterprise SOC 2/HIPAA deployments at scale.
| Plan | Monthly price |
|---|---|
| Starter | $50/mo |
| Professional | $200/mo |
| Enterprise | Custom |
Starter: $50/month flat plus usage, 1 active user, unlimited datasets/monitors/Sentinels, 30-day production traffic retention. Professional: $200/month per team plus usage, up to 5 active users, 1M indexed requests/month, 90-day retention, 3 hosted models included. Enterprise: custom pricing with continuous automatic model improvement, dedicated inference, on-prem/VPC deployment, SSO/SAML, and signed DPAs.
Maitai is a YC S24 control plane for production AI that helped lift Phonely's voice-agent accuracy from 81.5% to 99.2% across four model iterations, per a 2026 VentureBeat report. It indexes production traffic, deploys real-time Sentinels to intercept and correct bad model output, and automatically fine-tunes and distills task-specific models from live usage data.
Real-time Sentinels
Deploys guardrails that monitor every model response in production and intercept or auto-correct bad output before it reaches the end user.
Production Learning Loop
Indexes every request, trace, and outcome from live traffic into a structured, searchable log so evaluations and later training runs pull from real usage instead of synthetic tests.
Automated fine-tuning and distillation
Turns curated production traces into training data automatically, fine-tuning and distilling task-specific models so inference gets faster and cheaper without a manual ML pipeline.
Workflow and agent version control
Versions models, workflows, and agents as code with staged promotion from staging to production and one-click rollback between releases.
Maitai Mojo copilot
An in-Portal contextual assistant that can be asked to analyze runs or commanded to build datasets and kick off fine-tuning jobs directly, without writing new pipeline code.
Voice-agent reasoning engine
Powers voice agents with a 96ms p95 response time in Maitai's published case study, keeping multi-turn calls coherent while orchestrating tool calls like scheduling and payments.
Agent ACLs
Restricts sub-agents with access control lists so a parent router can hold full context while downstream workers only see the data explicitly allowed, protecting PII.
Most users expect these and they are not currently supported.
| Tool | Price from | Free plan | Runs on | Rating |
|---|---|---|---|---|
| $50/mo | No | Web | -- | |
| -- | Yes | Web | 3.0 | |
| Quote | No | Web | 3.5 | |
| -- | Yes | Web | 3.9 |
Proprietary
Starter is $50 a month flat plus usage, covering one user, unlimited datasets, monitors, and Sentinels, and 30-day traffic retention. Professional runs $200 a month per team plus usage, extending to five users, 1M indexed requests a month, 90-day retention, and three hosted models. Enterprise pricing is custom and adds continuous automatic model improvement, on-prem or VPC deployment, dedicated infrastructure, and signed DPAs.
Maitai has no $0 tier. Starter is the entry point, and the company says no credit card is needed to explore the Portal before committing. There is no separate trial period disclosed, so evaluation happens on the paid Starter plan itself rather than a free sandbox.
Crukx suits teams that want white-glove, contract-only enterprise observability with no self-serve option. Openlayer suits teams that want a free eval tier with built-in EU AI Act compliance tests before paying anything. SuperPenguin fits better if the priority is attributing AI spend to features and customers rather than correcting model output.
Both position themselves as production LLM control planes, but Crukx sells only custom, contract-based enterprise deals with no published pricing, while Maitai publishes self-serve plans a team can sign up for directly. Maitai also bundles automated fine-tuning and distillation into every plan; Crukx's public pages don't disclose whether fine-tuning is included. Teams that want to start without a sales call are better served by Maitai.
Drop the Python SDK, published as maitai-python on PyPI, into an existing app to start capturing traces, then log into the Portal to view indexed requests. From there, Maitai Mojo, the in-portal assistant, can build a first dataset and kick off fine-tuning without new pipeline code. Full model training still depends on how quickly enough live traffic accumulates.
Maitai is a control plane for production AI built by Christian DalSanto and Ian Hoegen, who founded the San Francisco company in 2024 and took it through Y Combinator's Summer 2024 batch on a standard $500K seed. The team is small, around six people as of 2026, and pitches the product as an alternative to hand-patching prompts every time a model drifts: instead, Maitai indexes every production request, trace, and outcome, evaluates it in real time, and feeds the results back into better models and workflows.
The mechanism is a three-stage loop the company calls Curate, Monitor, Improve. A drop-in SDK (the Python package is published as maitai-python on PyPI) captures traces from an existing application, the Portal evaluates that traffic live and applies guardrails called Sentinels that intercept and correct bad model output before it reaches a user, and the curated examples get turned into fine-tuned or distilled task-specific models over time. Version control extends to workflows and agents too, with staged promotion and rollback between releases, plus access control lists that keep sub-agents from seeing data a parent agent has not explicitly shared.
Maitai's published case studies point at regulated or latency-sensitive use cases: a customs-classification workflow that reasons over more than a million HS tariff codes, and a voice-agent deployment built with Groq and Phonely that the companies say lifted response accuracy from 81.5% to 99.2% across four model iterations, ahead of GPT-4o's 94.7% baseline, per a 2026 VentureBeat report. Buyers are ML and platform engineers at companies running AI agents where a wrong answer is expensive: contact centers, voice support, and back-office classification.
Pricing is self-serve up to a point: a Starter plan for individuals, a per-team Professional plan, and a custom Enterprise tier for on-prem or VPC deployment. The product ships as a web Portal plus SDK, with no desktop or mobile client.
Product facts on this page (pricing, plans, integrations, compliance) are read from Maitai's own pages and documentation. The verdict is Hokai's reading of them, not the vendor's claim.