01 / Build This Weekend
Waymo and Cruise spend millions a year labeling LiDAR by hand. Foundation models just hit expert-level accuracy on 3D point clouds.
What just became possible
A March 2026 arXiv paper validated foundation models for automated 3D LiDAR annotation using expert guidelines, hitting 70% manual labor reduction with no accuracy loss. Point Transformer V3 and MinkUNet fine-tuned on SemanticKITTI and Waymo Open Dataset provide the base. Cloud-deployable as a SaaS pipeline with human-in-the-loop validation for edge cases.
Why now
L4 AV companies (Waymo, Cruise, Aurora, Zoox) generate over 10,000 hours of raw sensor data per week. They spend 15-25% of $556B annual AV development budgets on data annotation. LiDAR annotation alone is a $1.2B to $1.8B/year market by 2026 across the industry. North America leads with $841M and Scale AI plus Appen own most of it on manual headcount.
What you'd build
A cloud-based SaaS that ingests raw LiDAR scans and outputs production-ready semantic segmentation labels using fine-tuned foundation models. Sell to L4 OEMs and Tier-1 suppliers. Pricing: per-scan, with SLA on accuracy and edge case handling. LiDAR engineer salaries: $76k to $240k for 700+ active US postings in 2026.
Who's already moving
Scale AI (manual + AI-assisted, dominant). iMerit (human-in-loop). Deepen AI (multi-sensor). SuperAnnotate (enterprise). Segments.ai and Taskmonk on the automation side. No fully automated, cloud-based SaaS exists for production-grade LiDAR semantic segmentation. The buyer wants automation but won't accept anything less than human-in-loop quality.
The gap
Edge case handling on diverse fleet data. Domain adaptation across geographies and seasons. Integration with existing labeling workflows. DARPA AI programs offer $500k to $5M. DHS C-UAS has $250M allocated for 2026. NSF, USDA, NIH have up to $2M per project. The deal: land one L4 OEM as anchor customer, build the edge case dataset, expand.
02 / AI Makes This Possible
You can now generate a synthetic LiDAR scan and its pixel-perfect segmentation mask at the same time. AV teams stop waiting for real-world edge cases.
What just became possible
A March 2025 arXiv paper showed joint generation of high-fidelity images and precise segmentation masks conditioned on text and spatial prompts. SynDiff-AD and similar diffusion architectures produce geometrically consistent paired training data. No manual correction. Rare scenarios (pedestrians in heavy rain, unusual vehicle configurations) generated on demand.
Why now
AI annotation market: $2.3B in 2024 to $28.5B by 2034 at 28.6% CAGR. Synthetic driving data: $1.14B to $9.17B at 24.7% CAGR. Automotive AI simulation and synthetic data overall: $1.03B to $29.15B at 39% CAGR. Perception software is 38.5% of autonomous driving software market. The supply-demand pull is real and growing.
What you'd build
An automated data synthesis platform that generates paired image-and-mask training data on demand. Sell to AV perception teams at Tier-1 suppliers and OEMs. Charge per million samples, with SLA on diversity and edge case coverage. The buyer is the perception lead trying to ship a new model release.
Who's already moving
Synthesis AI ($17M, synthetic people/environments). Applied Intuition ($250M raise, simulation for AV validation, 20+ AV developers). Wayve ($1.05B Series C, embodied AI). Parallel Domain (synthetic data). NVIDIA DRIVE Sim (neural rendering, 20+ AV developers). CARLA Simulator (open source). The whitespace: joint image-and-mask diffusion productized, not just simulation.
The gap
Photorealism with sensor-physics accuracy (LiDAR returns, camera noise). Diversity that captures the long tail. Validation that synthetic-trained models transfer to real-world performance. NEDO funds TIER IV in Japan. EU Horizon Europe has €100M+ for AV. NSF has $50M+ in active grants. The first to ship 10M paired samples that train models to AV-grade performance owns the segment.
03 / Deep Tech Bet
Federal agencies can't use cloud LLMs. Open-source 8B models just matched GPT-4 on cybersecurity tasks while running air-gapped.
What just became possible
A March 2026 arXiv paper showed open-source LLMs (Llama 3.1 8B, Qwen 2.5 7B, Cisco's Foundation-sec-8b) hitting GPT-4-comparable performance on cybersecurity benchmarks while running on local hardware. Quantized for RTX 4090 or A100. Ollama, vLLM, and OpenLLM Air-Gapped Edition handle the deployment. The capability gap closed.
Why now
AI in defense and security: $15.96B in 2026 with $13.4B in DoD AI spending alone, growing at 12.5% CAGR to $25.58B by 2030. IL5/IL6, CMMC, FedRAMP, and the new CMMC 2.0 framework prohibit cloud AI for classified data. Federal agencies are mandated to find air-gapped solutions. The procurement cycle is active right now.
What you'd build
A self-hosted, air-gapped security operations platform that runs specialized open-source LLMs locally. Threat analysis, incident response, log triage, no external API. Buyer: federal agencies and defense contractors handling Top Secret/SCI data. Hardware investment $800k to $1.2M per deployment, $300k integration. 18-month break-even versus cloud APIs.
Who's already moving
AirgapAI (2,800+ pre-built workflows, $697 perpetual license, SCIF/HIPAA/CMMC). RedSage (8B agentic cybersecurity, locally deployable). Cisco Foundation-sec-8b (open weight, permissive license). Palantir AIP (IL5/IL6, but $5M+ deployments). Microsoft Azure OpenAI disconnected containers. SentinelOne Prompt Security launched at RSAC 2026. Crowded but classified-tier work is narrow.
The gap
FedRAMP and IL6 certification (the slowest part). Specialized fine-tuning on cyber threat intel datasets. Kubernetes-based deployments that survive air-gap update cycles. DARPA I2O FY2026 BAA is open. NIST AI Manufacturing USA Institute has $70M over five years. AI/ML security engineers earn $200k to $280k+. The contract that lands is the one with the right clearance and the right deployment footprint.
Child malnutrition programs spend billions on imported peanut paste. A finger millet formulation matches Plumpy'Nut on outcomes and is made locally.
What just became possible
A November 2025 medRxiv paper validated Balamrutham+, a finger millet-based ready-to-use complementary food, showing improved nutrient bioavailability and gut microbiota modulation for moderate acute malnutrition. Clinical trial endpoints met. Shelf-stable, fortified, palatable. Locally producible at fraction of imported RUTF cost.
Why now
45.4 million children under five suffer acute malnutrition globally. Sub-Saharan Africa alone is a $2.1B RUTF/RUSF market in 2024. Total addressable: $3-5B annually across Sub-Saharan Africa and South Asia. UNICEF and WFP are explicitly shifting to local manufacturing to reduce supply chain risk on Plumpy'Nut.
What you'd build
A GMP-certified manufacturer of fortified finger millet ready-to-use complementary food. Process in-country to UNICEF/WHO technical specifications. Sell to UNICEF Supply Division, national health ministries, and NGO procurement teams. Or build the contract manufacturing infrastructure for local producers to scale.
Who's already moving
NutriBoost (Burkina Faso, operational, scaling in West Africa). Fountainhead Foods (India, $500k+ turnover). Balamrutham+ (clinical validation, government partnerships in India). Nestlé, Danone, Nutriset dominate imported peanut-based RUTFs but face supply chain and localization pressure. Local SMEs are the preferred partners for governments and NGOs.
The gap
GMP-certified infrastructure for local millet processors. Quality control to WHO/UNICEF specs (micronutrient levels, shelf-life in high humidity, palatability). National regulatory approvals (FSSAI in India, KEBS in Kenya). GAFSP ($2.44B mobilized since 2010), CFPCGP ($250k/year), WFP-DSM partnerships fund this. The first three GMP-certified African or South Asian millet RUTF facilities win the next decade of procurement.
05 / Watch This Space
SEC filings are still parsed by Python scripts and Kofax. An open-source extractor just hit near-perfect accuracy on formulas and complex layouts.
What just became possible
A September 2024 arXiv paper introduced open-source extraction tools that preserve formulas, table layouts, and reading order across complex PDFs with near-human accuracy. LlamaExtract, Qianfan-OCR end-to-end models, and layout-aware VLMs (LayoutLM, Donut) combined eliminate the pipeline-error problem. Spatial-semantic understanding in one model.
Why now
Intelligent document processing market: $14.16B in 2026 to $91B by 2034 at 26.2% CAGR. BFSI is 31.7% of global IDP, finance and accounting is 45.6% of all IDP spend. Quant funds process 10,000+ SEC filings daily. Current spend on manual data entry plus legacy OCR (Kofax, ABBYY) is in the hundreds of millions annually. ABBYY and Kofax are vulnerable to a focused open-source-first competitor.
What you'd build
An API-first document intelligence platform that converts complex 10-K/10-Q filings, technical reports, and financial statements into structured JSON. Preserve formulas and tables with high fidelity. Sell to quant funds, asset managers, and regulatory compliance teams. Pricing: per-page or per-filing with volume discounts.
Who's already moving
LlamaIndex (LlamaExtract, used for SEC filings). AlphaSense (agentic AI for financial research, used by 80% of top asset managers). Zillion (10-K/10-Q rapid analysis). ABBYY, Kofax, UiPath, Automation Anywhere are the incumbents. The whitespace: an API that handles the actual hard cases (multi-page tables, footnoted formulas) without per-document tuning.
The gap
Layout-aware VLMs that survive obscure PDF templates. Continuous fine-tuning on new filing formats. NSF TechAccess: AI-Ready America funds the research foundation. USDA-NIFA and DOE AI/Quantum have additional pots. The Bloomberg, FactSet, Refinitiv buyers expect enterprise SLAs. Land 3 top-20 hedge funds as reference customers and the segment follows.
See you next week.
- Theis