PAIR Systems

PAIR Systems

Autonomous IR for enterprise AI systems

We are building the infrastructure and optimization layer that lets every enterprise deploy a world-class, self-improving agentic memory system.

Lead with the company thesis, not the product SKU. “Autonomous IR” is the larger story. GoodMem is the first deployable product in that direction.

Why Now

Research convergence

• 2017: Zero shot neural retrieval was first demonstrated
• 2018-2019: multilingual retrieval became real
• 2020-2021: CLIP showed retrieval could span text and vision
• 2025-2026: universal embeddings and agent-driven optimization loops make self-tuning retrieval plausible

Market convergence

• RAG moved from demo to production
• Agentic systems make retrieval and memory core infrastructure
• Enterprises need governance, determinism, and deployment control
• Most organizations cannot staff a dedicated ML search team
• RAG and agentic memory are going mainstream, with no clear winner yet

This slide should explain why the technical and market timing now line up. Do not over-teach the research, but do make the arc legible.

The Problem

The bottleneck is no longer model access. It is retrieval quality.

What breaks

• Every enterprise agent is only as good as the context it retrieves
• Teams still stitch together memory, embeddings, rerankers, and inference by hand
• Retrieval quality is inconsistent and hard to improve
• Governance, auditability, and lifecycle controls are often afterthoughts

What enterprise buyers actually need

• World-class retrieval without hiring a scarce ML search engineer
• A deployable control plane for memory and retrieval
• Measurable self-improvement on the customer’s own corpus
• Governed, provider-neutral infrastructure teams can standardize on
• A path from text retrieval today to multilingual and multimodal retrieval over time

Keep this buyer-centric. The key move is shifting from “memory problem” to “retrieval engineering problem.”

Why This Team

Amin Ahmad, Founder and CEO

• Pioneered neural retrieval at Google Research in 2017
• Helped push retrieval into multilingual and cross-lingual systems
• Founder and original architect of Vectara
• Took Vectara from zero to launch, early enterprise traction, and a company that continues to scale
• Now building the next step: autonomous IR for enterprise systems

Forrest Sheng Bao, Co-founder and Head of ML

• 10+ years as a published ML researcher
• Former professor and later ML leader at Vectara
• Adds evaluation, optimization, and applied model depth

Selected Team

• Tom Diffenbach: ex-Google L6, senior systems
• Zaid Abdurehman: ex-Oracle, security and hardening
• Rogger Luo and Weisi Fan: PhD-level ML and optimization depth

This slide is early on purpose. Amin’s career arc is part of the product thesis, not generic resume material.

What GoodMem Is

The foundation for autonomous IR

Memory

• Managed memories
• Retrieval pipelines
• Persistent storage
• Data isolation model

Retrieval

• Retrieval pipelines
• Query logging
• Self improvement

Resource + Model Control

• Managed embedders/rerankers/LLMs
• PG persistence layer
• Local and cloud inference integration

Enterprise Control Plane

• Diagnostics and health
• Lifecycle controls
• Extension points
• API-first and self-hostable

GoodMem should be framed as a memory + retrieval + inference control plane for enterprise AI systems.

Becomes more valuable as PAIR’s optimization loops turn it into a self-improving retrieval system.

This is where GoodMem enters the story: as the deployable substrate, not as the whole company.

Autonomous Self-Improvement Is the Key

Goal: Automate the work of an expert ML search engineer

What Pair should automate

• Evaluate retrieval quality on a customer’s own corpus
• Fine-tune embedders and rerankers automatically
• Run closed-loop experiments, regression detection, and model selection
• Deliver expert-level IR improvement as software

Why it matters

• Off-the-shelf models leave performance on the table in specialized business settings.
• Our team has demonstrably beat OpenAI embedding models.
• ML engineers are a scarce resource.

This slide is now strategic, not optional.

Demos

• fly.io install and deploy
• Sailpoint KB Agent

Product Roadmap

Done

• Multilingual, text focused RAG system
• Model and infrastructure agnostic
• Enterprise hardening - FIPS, SLSA 3, CVE scanning
• Basic Cloud Fine Tuner

Automate

• Multimodal memory - Audio, Image, Video, Tabular, Screenshot
• Self-tuning retrieval that easily beats OpenAI embedding models.
• Measurable uplift in KPIs without a dedicated search-ML team

Others

• SIEM and auditability
• Enhanced evidence citation

Start where enterprises already have pain today. Expand toward a universal retrieval layer that improves itself over time.

This is the roadmap slide without turning into a feature checklist. The progression is substrate to automation to broader coverage.

Early Commercial Proof

$60K

Live recurring ARR today

$460K

ARR if current Incorta infrastructure partnership closes

$120K

Total historical revenue

Evidence that GoodMem is already selling

• Wanclouds validated paid demand ($35k)
• A Fortune 50 customer is live and paying through the Incorta OEM channel ($60k)
• Incorta has an unsigned $400K/year infrastructure partnership proposal in motion
• Agiletek, Slalom, and Bering support the emerging SI channel
• In early discussions with Sailpoint and ServiceUp

Stay precise. The point is not huge current ARR. It is that the deployable control-plane layer already monetizes, which supports the broader company thesis.

Business Model and GTM

Four revenue motions

Infrastructure partner

$300K-$400K+

Predictable annual platform license

OEM per-instance

$25K-$110K

Annual subscription per customer instance

SI channel

Enterprise and federal access

Direct enterprise

Higher ACV and tighter product feedback loops

Adoption flow

1. Enter

Through OEMs, SIs, or direct enterprise demand

→

2. Deploy

Customers buy infrastructure plus help getting it into production

→

3. Expand

Production wins drive renewals, more instances, and direct pull

• Forward-deployed engineers speed deployment and partner enablement
• We are not building a large services bench ahead of demand

This slide should now read structurally at a glance: four ways the company gets paid, a three-step adoption motion, and a feedback loop from deployments back into product advantage.

Market

Derived from Gartner’s 2026 AI software categories, using only the slices that map to GoodMem

$4.5B-$7.5B

Core TAM

$1B-$2.5B

Near-term SAM

$30M-$100M

3-5 year SOM

• Three Gartner 2026-spend categories: AI Platforms for DS/ML (31.3B), AI Application Development Platforms (8.4B), and AI Data (3.1B)
• TAM captures only the retrieval, memory, reranking, and inference-control share of those categories (5-10%; 25-35%; 25-50%, respectively).
• SAM narrows to self-hosted, provider-neutral, API-first buyers (25-35%)
• SOM assumes a low-single-digit 3-5 year share of that near-term market (3-4%), with autonomous optimization as the longer-term upside

Source model: Gartner worldwide AI spending forecast, January 15, 2026. Category anchors used in the internal TAM model: AI Platforms for DS/ML ($31.1B), AI Application Development Platforms ($8.4B), and AI Data ($3.1B).

Keep this to one slide. The category definition matters as much as the headline number, but now it also has to support the larger autonomous-IR thesis.

The Raise

• $10M for 30-36 months of Runway.
• Make GoodMem the default autonomous IR stack for enterprise AI

Use of funds

• Harden GoodMem as the enterprise control plane
• Build the autonomous optimization layer and benchmark proof
• Expand GTM across OEMs, SIs, and direct enterprise
• Target $250k/month burn max. Currently $140k.
  • Stay capital-efficient: no large sales org, no broad international expansion, no large services bench ahead of demand

24-month targets

• Reach 10+ production deployments and $2.5M-$3.5M ARR
• Expand Incorta to >$500K ARR
• Add 1+ additional OEM / platform partner
• Add 4-5 direct customers
• Prove repeatable self-tuning retrieval gains on customer-specific data (beating Open AI)

This slide should answer what the money buys: substrate hardening, autonomous optimization proof, and a diversified commercial engine.

The Vision

The Vision

Make self-improving enterprise memory and retrieval deployable everywhere

As agentic AI takes on real enterprise decisions, self-improving memory over multimodal, unstructured data becomes mandatory. PAIR Systems will be the foundational infrastructure for that transition.

End on the larger mission. The point is not only better infrastructure. It is making self-improving enterprise memory and retrieval infrastructure deployable as software.

Appendix: Competitive Framing

The field breaks into four practical groups

• Direct enterprise RAG / agent platforms
  • Vectara, Contextual AI, Credal, Ragie, Onyx, RAGFlow
• Memory-layer competitors
  • Mem0, Zep, Cognee
• Enterprise search / generative answer incumbents
  • Glean, Elastic, Coveo
• Hyperscaler alternatives
  • AWS, Google Cloud, Azure

Appendix: Competitive Framing

GoodMem competitive landscape

Appendix: Revenue Model Detail

Current revenue architecture

• Embedded infrastructure partnership
  • flat annual license for partner-controlled AI features
• OEM per-instance licensing
  • partner sells GoodMem-powered capability into end-customer accounts
• SI channel
  • surfaces GoodMem inside enterprise transformation projects
• Early services
  • acquisition path, not long-term business model

Current proof points

• Fortune 50 via Incorta: $60K/year, live and paying
• Incorta infrastructure partnership: $400K/year, unsigned
• Wanclouds: $30K, one-time historical revenue

Appendix: Market Sizing Assumptions

How the top-down model is built

Anchor category	2026 spend	GoodMem carve-out	Why it maps	Implied contribution
AI Platforms for DS/ML	$31.1B	5-10%	Retrieval, reranking, memory, and inference-control slice inside broader AI platforms	$1.6B-$3.1B
AI Application Development Platforms	$8.4B	25-50%	Agent / RAG application layer where retrieval and orchestration are central	$2.1B-$4.2B
AI Data	$3.1B	25-50%	AI data infrastructure touching memory, retrieval, and serving	$0.8B-$1.6B
Core TAM				$4.5B-$7.5B

Filters from TAM to SAM to SOM

• SAM = $1B-$2.5B Filter assumes roughly 25-35% of the core TAM values self-hosting, provider neutrality, API-first infrastructure, and controlled enterprise deployment
• SOM = $30M-$100M ARR Assumes a low-single-digit 3-4% share of SAM over 3-5 years for a differentiated infrastructure company
• Menlo cross-check Menlo’s 2025 enterprise GenAI spend breakdown shows about $1.5B in AI infrastructure; that supports the low end of the Gartner-derived range but does not independently confirm the high end
• Boundary condition This is a software TAM for the memory / retrieval / control-plane layer, not total AI spend, GPU spend, services, or model API usage

External anchors used in the internal market model: Gartner worldwide AI spending forecast and Menlo Ventures’ 2025 enterprise GenAI spend breakdown. Carve-out percentages and share assumptions are internal market-model judgments.

Appendix: Cloud Fine-Tuner on CRUMB

CRUMB task IDs translated into user-facing retrieval problems

CRUMB task	Plain-English task	Base (MRR)	Fine-tuned	Delta
clinical_trial	Clinical trial matching from patient histories	0.6333	0.7458	+0.1125
code_retrieval	Code solution retrieval for multi-constraint problems	0.3889	0.6207	+0.2318
legal_qa	State-specific legal statute retrieval	0.2316	0.2841	+0.0525
paper_retrieval	Scientific paper retrieval from multi-aspect criteria	0.4494	0.4512	+0.0018
set_operation	Set-based entity retrieval	0.2583	0.2628	+0.0045
stack_exchange	Reasoning-heavy community QA retrieval	0.2141	0.2886	+0.0745
theorem_retrieval	Mathematical theorem retrieval	0.3125	0.3266	+0.0141
tip_of_the_tongue	Vague movie / TV retrieval from remembered details	0.0674	0.1387	+0.0713

What this says

• The cloud fine-tuner is already producing fully automated gains on several genuinely hard retrieval categories
• Biggest improvements here are code retrieval, clinical trial matching, StackExchange QA, and tip-of-the-tongue retrieval
• Small deltas on paper retrieval, set operations, and theorem retrieval suggest where more task-specific tuning is still needed

Task labels adapted from Killingback and Zamani, Benchmarking Information Retrieval Models on Complex Retrieval Tasks (CRUMB, arXiv:2509.07253), plus the public CRUMB benchmark repository. Scores shown here are from Pair’s internal cloud fine-tuner output.