Self-improving AI infra that gets cheaper with use

Occam is building the first self-improving AI infrastructure platform that automatically gets both cheaper and smarter with every query processed.

We start with cascade routing that cuts LLM costs by 90 percent - but unlike static routers , our system continuously learns from usage patterns, to compound savings beyond 90 percent over time. Every query makes the infrastructure smarter: smaller models automatically learn from larger models' outputs, query patterns train specialized models, and routing decisions optimize themselves. A startup spending $30k/month on GPT-4.5 drops to $3k/month while maintaining 95%+ accuracy. Improvements in latency response time, model accuracy, and cost savings improve automatically without any configuration.

With Routing as our wedge, we're looking to build the complete AI infrastructure platform where savings and precision compound in value over time. As we process queries, we're learning which patterns are unique to each organization, automatically training custom models for common tasks, and building intelligence that becomes a competitive moat. Future capabilities will include automatic model fine-tuning, query-specific model generation, and predictive optimization that anticipates needs before they arise.

Seattle, WA / San Francisco, CA

N/A (Was not a question on this application)

We have met each other as colleagues at PNC Bank's summer internship 2023. Since then, we've been close peers working alongside each other at hackathons and competitions.

2 Weeks:

Week 1: Dialed in on problem-space, dove into research for solution, and finalized our initial product offering
Week 2: Refining our long-term vision beyond our initial wedge into the market, and coding the API

Current Progress:
-> API fully functional
-> Platform is working locally but has yet to be hosted

Both of us have been prototyping and testing ideas for the past month together with multiple different products and markets, but have worked on this for the past two weeks.

August 31st

N/A

N/A

We discovered this problem through our combined experiences at enterprises watching them burn $50k+/month on AI with many projects reporting zero ROI - not because AI doesn't work, but because the economics are broken. Enterprise AI spending doubled in 2025, but companies are trapped between expensive accuracy and cheap failure.

Our domain expertise is complementary:
Ken's work at Google on distributed AI systems (Ray on Kubernetes, TPU/GPU optimization) is critical for scaling small/proprietary SLMs - each enterprise needs custom GPU optimization and hosting, which is exactly what he builds at scale.
I bring a unique blend of data science expertise, and pain-point identification from my roles as a Angel VC analyst + PM @ startups and F500s. I’ve honed in my product instincts from shipping across domains and founding EcoTrace which taught me to identify which technical innovations actually drive value.

We picked this idea because the market is screaming for this: Menlo Ventures reports LLM API costs doubled to $8.4B in 2024 while McKinsey shows 80% of enterprises see no ROI. Every company spending >$10k/month on LLMs needs this yesterday. We validated the technology by implementing the academic research ourselves: reproducing the "Online Cascade Learning" results (UT Austin/Rice, 2024) across multiple tasks and systems, achieving the promised 90% and compounding cost reduction. We realized this could be productized as self-improving infrastructure that compounds in value over time through reducing latency, costs, and inaccuracy.

Here's what we think everyone's missing: infrastructure should get smarter and cheaper the more you use it, not stay frozen in time. Our cascades learn from every query - smaller models automatically train on larger model outputs, routing decisions optimize themselves, and your system builds organizational intelligence that becomes your moat.

The technical moat is implementing online cascade learning - it requires deep expertise in episodic MDPs, confidence calibration, and distributed GPU optimization that took researchers years to crack. Every query makes the system smarter, cheaper, and faster. While AWS gets more expensive as you scale, we get cheaper.

Our vision is beyond just undercutting existing LLM router solutions on price. We're building the economic infrastructure that democratizes proprietary AI - making it possible for any company to build and scale their own intelligent systems without burning millions. That's a much bigger prize than being the best router.

Here's our anticipated revenue model:

Usage-based: We'll charge 5% of inference costs saved
If a customer saves $30k/month → pays us $1.5k/month, because the high ROI makes purchasing decision easy

Enterprise platform: $10-50k/month for custom model training, compliance, SLAs

Market Opportunity:
Current: $8.4B spent on LLM APIs in 2024, with LLM market growing 33% CAGR to $36B by 2030
Addressable: 50,000+ companies spending >$10k/month on LLMs
Conservative Potential: 10% market share of API costs = $3.6B+ in processed volume = $70-100M annual revenue

N/A (Not a question on our application cycle)

No

Both

N/A (Not a question on our application cycle)

Misty: Agent for cloud infrastructure maintenance
- Allows companies which small/non scaled deployment strategies to deploy their applications on cloud providers
- Bridges human language to cloud infrastructure actions like kubectl, gcloud, etc. Ex: “Hey Misty can you set up a kubernetes cluster hosted on GKE make sure it scalable to the entire U.S. and within our budget of $1000 for the next 6 months”

Sophon: AI Governance/Compliance System
- Chain-of-thought guardrails with perceived agent actions vs actual agent actions
- Human-in-the-loop systems and security access checkpoints
- Industry specific tailoring for compliance in domain-specific workflows (Eg. Healthcare, Banking)

N/A (Not a question on our application cycle)

This is our second time founding together after EcoTrace, which we built at Pitt and sold to our university. That experience taught us two critical lessons: pick massive markets (not just university procurement) and solve real pain points (not theoretical problems).

Coming from Pitt shaped our approach. Unlike the elitist/grifting (no offense) culture of building for prestige or the next cool thing, Pitt taught us to be scrappy no-BS builders who solve unglamorous but critical problems. That's why we're tackling enterprise AI costs, not building another ChatGPT wrapper.

Three factors convinced us YC is essential for Occam:
- We have a technical head start: Our approach implements academic research on online cascade learning that achieves 90% cost reduction while improving over time. YC's deep technical network can help us maintain this moat as we scale.
- The pain point is LOUD: Enterprises are burning $20k+/month on LLMs with the majority of investments seeing zero ROI. OpenRouter just raised at $100M+ valuation with a static solution. We're building the self-improving version that will obsolete them, and we need YC's velocity to capture this market before others catch on.
- Network effects require perfect execution: Every query makes our system smarter for all customers. YC has the playbook for these compounding advantages (Figma, Notion) that turn early leads into permanent moats.

Kenny and I are both technical founders who've intentionally built complementary skills - I've done PM work and VC to understand market dynamics, while Kenny went deep on distributed systems at Google. We're ready to build something significant early in our careers, and Occam is the right vehicle at the perfect time.

We've known about YC since high school - Kenny through CS communities, me through debate circuits where both of my competition partners ended up becoming cracked CS students. But what shifted us from admirers to applicants was our time at Pitt, where we learned to identify authentic pain points rather than chase hype cycles. Our most impactful features in the first product we launched together wasn't the novel computer vision architecture we deployed, but rather the simple phone automations that saved universities $13,000/year. Studying how YC companies like Docker and ScaleAI built developer infrastructure convinced us that YC uniquely understands how to scale technical infrastructure that becomes essential plumbing for the entire ecosystem. When we realized Occam could be the cost optimization layer that every AI company needs, and that our research-backed approach gives us a real technical moat, YC became the obvious choice.