# Softmax Data

> Boutique AI engineering firm (est. 2019). We build production-grade AI systems for PE-backed and mid-market companies — AI agents, document AI, workflow automation, data foundations, and custom model fine-tuning. 150+ production AI systems shipped. $72M+ in measured business value created.

## What we do

Softmax Data is a senior AI engineering team that designs, builds, and deploys custom AI:

- AI product features (customer-facing) embedded inside SaaS products
- Internal AI apps for teams (automation, reporting, decision support, agentic analytics)

We do not "install AI tools." We build and ship systems into production — with guardrails, monitoring, audit trails, and human-in-the-loop review where it matters.

## Who we serve (ICP)

- Private equity-backed portfolio companies needing AI to drive EBITDA and growth
- US/Canada software companies (~$50–150M ARR) shipping AI features into their products
- Marketing and sales agencies (50+ employees) automating labor-heavy reporting and ops
- Document-heavy industries: mortgage, real estate, insurance, healthcare, government, legal

## Services

### Document AI
Automate classification, extraction, and verification on messy scanned/rasterized PDFs. Computer vision preprocessing, multimodal models, human-in-the-loop review, continuous improvement via feedback loops.
URL: https://softmaxdata.com/services/document-ai

### AI Agents
Task-specific agents that analyze multi-source data with citations, audit logs, and controls. RAG-based retrieval, guardrails, monitoring, and integration with existing systems.
URL: https://softmaxdata.com/services/ai-agents

### Workflow Automation
Production workflows with human approval, audit logs, rollback, and monitoring. Automate repeatable processes while keeping humans in control. Scheduled runs, fallbacks, alerts.
URL: https://softmaxdata.com/services/workflow-automation

### Data Foundations (Databricks)
Lakehouse pipelines with data quality, governance, lineage, observability. Build the data infrastructure that makes AI reliable.
URL: https://softmaxdata.com/services/data-foundations

### Custom Models & Fine-Tuning
Fine-tune LLMs and build custom ML models when prompting alone isn't enough. Data preparation, training, evaluation, deployment.
URL: https://softmaxdata.com/services/custom-models

### AI Strategy & Use Case Mapping
Identify where AI can have the highest impact — mapping opportunities across operations, product, and customer experience, and prioritizing by ROI and feasibility.

## Open Source

### Engram — Context Database for AI Agents
Brain-inspired, portable context database for AI agents. Open-source (MIT license). Supports persistent memory, multi-agent context sharing, and reconsolidation-based learning.
GitHub: https://github.com/softmaxdata/engram
Website: https://engram.so
URL: https://softmaxdata.com/engram

## Proof (success stories)

- LauraMac (Custom Model / Document AI): 40 min → 38 sec processing time; 95% precision; 80% operating cost reduction. URL: /success-stories/lauramac
- Rocket Mortgage Canada (Document AI): <2 min per package; 90% less manual effort; 100% statements normalized. URL: /success-stories/rocket-mortgage
- Clio (Entity Resolution / GTM): 417% qualified opportunities; $6M new ARR. URL: /success-stories/clio
- Absolute Results (Predictive Models): 54% close rate; 210K+ vehicle sales. URL: /success-stories/absolute-results
- Flywheel (AI Agents): 60 hrs/week saved; 95.8% less time per report. URL: /success-stories/flywheel
- Predictable Revenue (Automation): ~20 min saved/meeting; ~10+ hrs/week. URL: /success-stories/predictable-revenue
- Enterprise Call Center (Voice AI): Real-time voice analysis; $150K annual savings; 8% productivity increase. URL: /success-stories/enterprise-call-center
- Premium Retailer (Agentic Analytics + Lakehouse): Avoided $3.5M loss; 17% inventory cost reduction; 8% sales increase. URL: /success-stories/retail-analytics

## Partnerships

- AWS Machine Learning Services Partner
- Databricks Partner
- Funnel.io Partner

## How we work

- Start with a 30-minute discovery call to understand the problem
- 5-day paid AI Prototype Sprint to demonstrate feasibility and results
- Typical engagements: 4 weeks to 6 months
- Milestone-based billing with bi-weekly demos
- ROI visible within 3 weeks

## Blog — AI Thought Leadership

Published at https://softmaxdata.com/blog — new posts every 2 days; site updated weekly.

We regularly analyze arXiv papers in depth — breaking down complex research (GRPO, PPO, DAPO, CL-Bench, ACE/ICLR) into plain language that any software engineer can follow. No PhD required.

Topics we cover extensively:
- arXiv paper deep dives — cutting-edge AI research explained in simple, practical language for engineers
- AI Agents & agentic design patterns (multi-agent architectures, agent swarms, tool use, MCP)
- LLM fine-tuning & reinforcement learning (GRPO, DAPO, PPO, LoRA, Unsloth, DeepSeek, Kimi — hands-on tutorials with code)
- Document AI & OCR (Textract, DeepSeek-OCR, production extraction pipelines)
- RAG & agentic RAG (vector databases, context engineering, retrieval strategies)
- Future of SaaS & AI's impact on software businesses (valuation, IP, vibe coding, build vs. buy)
- AI infrastructure (specialized chips, cost optimization, serverless inference)
- Agentic frameworks compared (LangGraph, CrewAI, AG2, OpenAI Agents SDK)
- AI strategy for PE-backed and mid-market companies (readiness, ROI, data quality)
- Prompt engineering, guardrails, evaluation metrics, and production best practices

### Recent posts (2026)
- From PPO to GRPO to DAPO: Understanding RL for LLMs and Every Training Parameter Explained (Mar 2026): https://softmaxdata.com/blog/from-ppo-to-grpo-to-dapo-understanding-rl-for-llms-and-every-training-parameter-explained/
- We Open-Sourced Engram: A Brain-Inspired Context Database for AI Agents (Mar 2026): https://softmaxdata.com/blog/we-open-sourced-engram-a-brain-inspired-context-database-for-ai-agents/
- Is LLM's Context Window a Vanity Metric? (Mar 2026): https://softmaxdata.com/blog/is-llms-context-window-is-a-vanity-metric/
- How to Tune Your Own LLM with GRPO, Common Crawl and Unsloth (Mar 2026): https://softmaxdata.com/blog/how-to-tune-your-own-llm-with-grpo-common-crawl-and-unsloth/
- In the Era of Vibe Coding, Is Code Really a Valuable IP at Exit? (Mar 2026): https://softmaxdata.com/blog/in-the-era-of-vibe-coding-is-code-really-a-valuable-ip-when-exit/
- Specialized AI Chips and How They Will Change Our Industry (Mar 2026): https://softmaxdata.com/blog/specialized-integrated-ai-chips-and-how-it-will-change-our-industry/
- The Future of Private, Small, Specialized AI Models — Why They May Be the Real Moat (Mar 2026): https://softmaxdata.com/blog/the-future-of-private-small-specialized-ai-models-and-why-they-may-be-the-real-moat/
- Definitive Guide to Agentic Frameworks in 2026: LangGraph, CrewAI, AG2, OpenAI and More (Feb 2026): https://softmaxdata.com/blog/definitive-guide-to-agentic-frameworks-in-2026-langgraph-crewai-ag2-openai-and-more/
- Agentic Design Patterns — Technical Deep Dive (Feb 2026): https://softmaxdata.com/blog/agentic-design-pattern-comparisoneveryones-talking-about-ai-agents-right-now-and-honestly-its-getting-hard-to-keep-up-every-week-theres-a-new-framework-a-new-sdk-a-new-revolutionary-app/
- CNN vs Transformer Model Difference in Image Processing (Feb 2026): https://softmaxdata.com/blog/cnn-vs-transformer-model-difference-in-image-processing/
- Context Matters: The Biggest Lesson from ACE at ICLR 2026 (Feb 2026): https://softmaxdata.com/blog/the-biggest-lesson-from-ace-iclr-2026-the-power-of-agentic-engineering/
- AI Agents, RAG, MCP, Workflow Automation: What They Actually Mean (Feb 2026): https://softmaxdata.com/blog/beats-with-many-heads-ai-agents-rag-mcp-workflow-automation/
- Agent Swarm vs Anthropic Workflows vs LangGraph: Which Multi-Agent Architecture? (Feb 2026): https://softmaxdata.com/blog/agent-architectures-compared/
- How to Fine-Tune Kimi K2.5 on Your Local Machine (Feb 2026): https://softmaxdata.com/blog/how-to-fine-tune-kimi-k2-5-on-your-local-machine-a-practical-guide/
- How to Fine-Tune DeepSeek OCR V2 on Your Own PDFs (Feb 2026): https://softmaxdata.com/blog/how-to-fine-tune-deepseek-ocr-v2-on-your-own-pdfs-from-install-to-inference/
- Why DeepSeek-OCR 2 Could Change Document AI Forever (Feb 2026): https://softmaxdata.com/blog/why-deepseek-ocr-2-could-change-document-ai-forever/
- The Three Layers of Useful AI: Agents, MCP, and SKILL.md (Jan 2026): https://softmaxdata.com/blog/ai-agents-skills-and-mcp-what-business-leaders-actually-need-to-know-2/
- SaaS at a Junction Point: What We Learned Building AI in 2025 (Dec 2025): https://softmaxdata.com/blog/saas-at-a-junction-point-what-we-learned-building-ai-in-2025/

### Technical deep dives (earlier)
- Making Serverless Inference on SageMaker for HuggingFace Models (Feb 2024): https://softmaxdata.com/blog/using-custom-huggingface-models-on-aws-sagemaker-severless-inference/
- Keras LSTM Source Code Line-by-Line Explained (Apr 2020): https://softmaxdata.com/blog/keras-lstm/
- Using ML to Understand Inside Sales Emails (Feb 2020): https://softmaxdata.com/blog/what-are-inside-sales-emails/
- Applying ML in Sales Enablement Part 1-3 (2019-2020): https://softmaxdata.com/blog/applying-machine-learning-in-sales-enablement-and-sales-operations-part-1/
- CRM Cleaning with Salesforce SOAP API (Aug 2019): https://softmaxdata.com/blog/crm-cleaning-merging-salesforce-objects-using-soap-api-with-a-python-focus/

## AI Learning Resources

Plain-English guides to AI concepts for non-technical audiences:
URL: https://softmaxdata.com/learn

Topics covered: AI Agents, RAG, MCP, Workflow Automation, Document AI, Fine-Tuning, Tokens, Context Windows, Agentic AI vs Generative AI, Self-Evolving Agents, Agent2Agent Protocol (A2A), Prompt Engineering, Context Engineering, Harness Engineering, and more.

## Start here

- Book a discovery call: https://softmaxdata.com/contact
- Services overview: https://softmaxdata.com/services
- Success stories: https://softmaxdata.com/success-stories
- FAQ: https://softmaxdata.com/faq
- Try document extraction demo: https://softmaxdata.com/demo/document-ai
- AI opportunity assessment: https://softmaxdata.com/demo/ai-assessment

## Contact

Jia Chen, Founder & CEO
Email: jia@softmaxdata.com
Website: https://softmaxdata.com
Blog: https://softmaxdata.com/blog