> AI Deployment Company

Production Deployments

Every production deployment our FDEs have shipped, with the problem, what we deployed, why it mattered technically, and the environment it runs in. Clients stay unnamed; the systems speak for themselves.

> 11 deployments across 32 domains. Expand any entry for the technical details.

◆ IRB-Approved Voice AI for Clinical Trials

Approved by two major IRBs

Healthcare · Voice AI · Regulated Deployment

Voice-native agents approved by two of the world's largest IRBs, deployed inside clinical trials for prescreening, informed consent, and scheduling.

Technical details

Problem: Clinical trial prescreening, informed consent, and participant scheduling all run under IRB oversight. The workflow called for voice-native agents that could hold a real conversation with participants, not chatbots.
What we deployed: Three voice agents embedded as platform-native workflows. Prescreening qualifies inbound candidates and routes them forward. Informed consent walks enrolled participants through the consent process by voice, answers questions in real time, and checks comprehension. Scheduling handles visit booking, reminders, and changes.
Technical significance: Among the first AI voice agents approved by two of the world's largest IRBs. Each workflow escalates to a human for clinical questions, and the agents are built into the trial platform rather than bolted on as widgets.
Production environment: Live inside a clinical trial platform, under IRB oversight, with human escalation paths for out-of-scope and clinical questions.
Outcome: Approved by two of the world's largest IRBs and running in production across all three trial workflows.

◆ Contact Intelligence & Sales Enablement Platform

Millions of enrichment requests

Deep Research Agents · Multi-Model Routing · VC-Backed Startup

AI engine for a venture-backed sales enablement platform: 30+ extraction agents behind a production API serving millions of enrichment requests.

Technical details

Problem: A venture-backed sales enablement platform needed contact enrichment, resume parsing, risk scoring, and fake profile detection at production scale.
What we deployed: The platform's AI engine: 30+ specialized extraction agents behind a production API, with multi-model routing that sends fast inference to Gemini Flash and complex reasoning to Claude.
Technical significance: Multi-model routing by task type keeps cost and latency in check while reserving the stronger model for the requests that need it. The agents run as a dependency of the product, not as a side tool.
Production environment: A production API inside a venture-backed SaaS platform, serving the product's live enrichment traffic.
Outcome: Serving millions of enrichment requests in production.

◆ Post-Training Evaluation for Coding LLMs

Eval loops at one of the largest LLM providers

LLM Infrastructure · Eval Harnesses · A/B Testing

Automated post-training eval loops and A/B test harnesses for agentic coding models at one of the largest LLM providers.

Technical details

Problem: One of the largest LLM providers needed to know whether each new agentic coding model was actually better than the last, and needed to know it fast, between training runs.
What we deployed: Automated post-training evaluation loops and test harnesses that A/B test coding models against real-world programming tasks at scale.
Technical significance: Evaluating agentic coding models means running them against real tasks, not benchmarks alone. The harnesses return rapid, comparable feedback on model quality while the models are still being trained.
Production environment: Post-training infrastructure at one of the largest LLM providers, in the loop between training runs.

◆ AI Gateway with Eval-Gated Model Routing

Running as a provider's production API

AI Infrastructure · Model Gateway · Inference

One endpoint for OpenAI-, Anthropic-, and Gemini-compatible SDKs, with evals built from real traffic deciding which model serves each request — running as an inference provider's production API.

Technical details

Problem: An AI inference provider needed one endpoint that any OpenAI, Anthropic, or Gemini SDK could call, with model selection decided by evidence rather than by hand.
What we deployed: A Go proxy with OpenAI-, Anthropic Messages-, and Gemini-compatible ingress; request-type classification; evals constructed from real production traffic; eval-gated routing that pins each application to the lowest-cost model meeting its quality and latency bar; a customer dashboard for API keys, traces, usage, and spend; and the operator tooling to onboard managed accounts.
Technical significance: Routing decisions are gated by evals built from the application's own traffic, so model changes are driven by measured quality and latency rather than by hand-picked defaults. Compatibility with three SDK families means applications adopt it without rewriting their clients.
Production environment: Runs in production as the provider's advertised API, with customer-facing and operator-facing tooling around it.
Outcome: In production as the inference provider's advertised API.

◆ Fully Agentic Back Office

Zero human operators for standard operations

SaaS Operations · Autonomous Agents · YC-Backed Company

Fully autonomous back office deployed for a YC-backed mobile app marketplace: support, billing, refunds, account management, zero humans in the loop.

Technical details

Problem: A YC-backed mobile application marketplace wanted its back office to run itself: customer support, inbound queries, account management, subscription billing, refunds, and FAQ resolution.
What we deployed: A fully autonomous back office that handles customer support tickets, inbound queries, user account management, subscription billing, refunds, and FAQ resolution.
Technical significance: Agents here take real actions with real consequences, including billing and refunds, on live customer accounts. The system had to be trusted to operate without a person reviewing each step.
Production environment: In production inside the company's operations, integrated with its support, billing, and account systems.
Outcome: Runs with zero human operators in the loop for standard operations.

◆ Self-Improving Agentic Software Factory

Coding Agents · Compound Engineering · Multi-Agent Orchestration

Closed-loop software factory shipped for a venture-backed startup, where humans and coding agents deliver production code through one pipeline.

Technical details

Problem: A venture-backed startup wanted humans and coding agents (Claude Code, Codex, Cursor, GitHub-native, autonomous loops) shipping production code through one pipeline instead of parallel, unreviewed workflows.
What we deployed: A single closed loop with multi-worktree parallel orchestration, cross-model review with disagreement gating, scoped agent identities, and branch-per-PR databases. A post-merge compound workflow turns each shipped change into reusable skills and playbooks.
Technical significance: Every bug becomes a regression test, every incident a runbook, every repeated agent mistake an updated AGENTS.md, skill, or eval. The pipeline improves itself with each change that ships.
Production environment: The startup's own engineering workflow and repositories, recently shipped and owned by the team.

◆ Academic Credential Verification at Scale

Deep Research Agents · ORCID · Scholar APIs

Deep research agents verifying PhD credentials across ORCID, Google Scholar, Semantic Scholar, and CrossRef for the largest data labeling platform.

Technical details

Problem: The largest data labeling platform had to verify PhD credentials and research publications at scale, across claims that span conference presentations, publication records, fields of study, and institutional affiliations.
What we deployed: Specialized deep research agents that cross-reference ORCID, Google Scholar, Semantic Scholar, CrossRef, and university databases to validate each academic claim.
Technical significance: Verification has to reconcile conflicting and incomplete records across several scholarly sources, and do it for every candidate rather than a sampled few.
Production environment: Runs credential verification for the platform at scale.

◆ Portfolio Intelligence for VC & PE Firms

RAG · Fund Operations · Hybrid Search

Portfolio intelligence, reverse portfolio search, and automated fund accounting built for several of the largest VC and PE firms.

Technical details

Problem: Several of the largest venture capital and private equity firms run fund operations on deal flow that arrives unstructured: emails, decks, PDFs, and notes.
What we deployed: Portfolio intelligence platforms built on RAG and hybrid search: reverse portfolio search engines, automated fund accounting workflows, and large-scale ingestion pipelines that turn deal flow into structured, searchable knowledge bases.
Technical significance: Hybrid vector and text search over the firm's own institutional knowledge, with ingestion pipelines that keep it structured as new deal flow arrives.
Production environment: Fund operations at several of the largest VC and PE firms.

◆ VC Due Diligence & Deal Flow Automation

Weeks to hours

Workflow Automation · Research Agents · Accelerator Programs

Deal flow triage, founder research, market sizing, and competitive analysis for accelerator and VC programs — weeks of due diligence reduced to hours.

Technical details

Problem: Multiple well-known accelerator and VC programs were spending weeks on manual due diligence: intake triage, founder background research, market sizing, and competitive landscape analysis.
What we deployed: Research agents that automate each of those diligence steps for the programs.
Technical significance: Each diligence step is a different research task with different sources. The agents are specialized per step and chained into one pipeline per deal.
Production environment: In use across multiple accelerator and VC programs' deal flow.
Outcome: The same diligence that took weeks now takes hours.

◆ Prediction Market Intelligence & Trading

Classical ML · Multi-Agent Research · Autonomous Trading

Multi-exchange prediction market platform: arbitrage detection across five exchanges, LLM research pipeline, Kelly Criterion trade execution.

Technical details

Problem: Cross-exchange arbitrage in prediction markets means matching equivalent contracts across venues and sizing every position correctly.
What we deployed: An end-to-end platform that aggregates data across five major exchanges, detects cross-platform arbitrage via fuzzy matching, and executes trades using Kelly Criterion sizing. A multi-stage AI research pipeline (Gemini for research, Claude for synthesis) feeds probabilistic confidence scores into classical ML trading models.
Technical significance: LLM-driven analysis is an input to traditional quantitative methods rather than a replacement for them. The hybrid design keeps position sizing on classical ML.
Production environment: Runs autonomously from data aggregation through trade execution across five exchanges.

◆ Autonomous AI Video Production

Content Generation · Video Rendering · Multi-Platform Publishing

Autonomous video pipeline — scripts, TTS, B-roll, data viz, FFmpeg composition — publishing to TikTok, X, and other platforms with no human in the rendering loop.

Technical details

Problem: Producing video from research through publishing normally keeps a human somewhere in the rendering loop.
What we deployed: A fully autonomous production pipeline: AI-generated scripts, local TTS, AI-generated B-roll, animated data visualizations, subtitle burning, and section composition via FFmpeg, publishing to TikTok, X, and other platforms.
Technical significance: Every stage from research to a rendered, subtitled, multi-platform video runs without a person in the loop.
Production environment: Runs autonomously, research to publish, with no human in the rendering loop.

What should we deploy?

If you have an important AI initiative that needs to move from idea, prototype, or stalled project into production, we'd like to hear about it.