2.5 trillion+ PDFs globally and counting. 90%+ aren't accessible to the blind. Manual remediation costs $5-25 per page and takes 15-30 minutes each. The April 2026 ADA Title II deadline is coming, and the math doesn't work.

Current approaches are semi-automated tools that still require 65% manual work and months of training. Or outsourced services with six-figure quotes and multi-week turnarounds. Neither scales.

What we built: Accessibility on Demand™ (or AoD™) is an automated platform that delivers remediated PDFs. 100% automation. 97-99% compliance scores validated by third-party checkers. 25,000 pages in 26 minutes for $6,000 (against quotes of $250,000 and four days.)

We picked this problem deliberately. It's not another productivity widget. It's equal access to information—something that matters—and a process that's been done the same manual way for decades. The kind of problem where thinking differently about the solution creates something 10X better, faster, and cheaper.

Accessibility on Demand™ is the first product from Netra Labs. It won't be the last.

If you're an IT leader staring at a remediation backlog and a deadline, or a partner looking for a solution that actually works at scale, would love to connect. 

Robert James Booth

robert.booth@netralabs.ai

Robert James Booth

Detailed: https://abiaryan.com/ | Technical Blog: https://modelcraft.substack.com/

Abi Aryan is the founder of Abide AI (www.abideai.com) and a machine learning research engineer with nearly a decade of experience building production-level ML systems. A mathematician by training, she previously served as a visiting research scholar at the Cognitive Systems Lab at UCLA, under Dr. Judea Pearl, where she focused on developing intelligent agents.

Abi has authored research papers in AutoML, multi-agent systems, and large language models, and actively reviews for leading research conferences and workshops, including NeurIPS, ACL (Association for Computational Linguistics), EMNLP (Empirical Methods in Natural Language Processing), and AABI (Advances in Approximate Bayesian Inference).

She is currently advancing research in three key areas:
	•	Reflective intelligence for AI agents
	•	Distributed self-healing protocols for multi-agent systems
	•	GPU engineering for very large-scale AI systems

Abi ︎

Companies learn this the painful way. Small vs. Large Models require fundamentally different system design skills and most existing infra engineers can't do both.

Once you cross ~70B parameters your work flips from inference optimization problem to a distributed resource management problem (also, known as the nightmare)

Here's the big differences when designing systems for them:

1️⃣ The biggest breaker is the Memory Regime Shift 
Small Models ≤30B have Weights ~60GB so FP16 fits 1-2 GPUs. The KV cache negligible and replication is cheap. 
For large models aka 100B-400B+, weights need to be sharded across 8+ H100s. KV cache dominates at long contexts. For example: 405B @ 128k context means ~400-500GB KV per request (FP16). One long session pins more memory than weights! 

Dynamic allocation, fragmentation, eviction means it's like running a distributed database, not just model serving

2️⃣ KV Cache turns inference stateful
For small models, ephemeral KV (~few GB total). Stateless batching shines. 
For large models, KV can exceed weights per request. 
@128k tokens: hundreds of GB pinned. And you run into issues like - head-of-line blocking, "one bad request" starves cluster, fairness explodes.

The modern fix for this is PagedAttention (vLLM) which reduces fragmentation 20-50%, but memory bandwidth (not FLOPs) still bottlenecks thus compute idles waiting on HBM.

3️⃣ Parallelism Overhead vs. Utilization 
For small models, it's near-100% GPU util easy. 
But for large models, Tensor + pipeline + (MoE) expert parallelism mandatory. 

And, stragglers kill latency. The reality is 60-70% utilization is "good" at scale (per reports on vLLM/TensorRT-LLM). If you trade efficiency for feasibility, interconnect saturation (NVLink 900GB/s) limits scaling.

4️⃣ Failure Modes do Flip 
For small models, restarts are cheap. 
But for large models, rehydration costs minutes (weights + KV). Partial failures strand GBs. 

So you need, fault isolation, graceful degradation to reduce retries.

5️⃣ Throughput vs. Latency Tradeoff Invert 
For small models, aggressive batching → max tokens/sec. 
But for large models, batching spikes memory → tail latency explodes (SLA violations). 

So the solution is to intentionally under-batch for predictable p99 latency.

6️⃣ Hardware Coupling becomes Inevitable 
For small models, you have portable abstractions. 
But for large models, co-design becomes mandatory. 

Think HBM capacity, NVLink topology, NUMA dictate architecture.

7️⃣ And finally, the Economics Are Non-Linear 
For small models, Cost ~ tokens. 
But for large models, fixed costs dominate (idle memory). Long context kills margins. Think memory-minutes, reserved tiers.

For small models, you need ML engineers with CUDA expertise.
For large models, you need distributed systems + compiler experts.

Takeaway: The game in 2026 is now learning how to master distributed systems, CUDA at the minimum. Where are you on that journey? Lets talk in the comments down below.

Abi Aryan ☯︎

Founder: Abide AI | Author: LLMOps & GPU Engineering | Making AI Systems go brrrr...

taplio

The Battle

Robert James Booth

Abi ︎

Followers

Posts

Engagements

Likes

Performance ScoreAbi ︎'s LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Robert James Booth

Abi ︎

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Abi ︎'s LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreAbi ︎'s LinkedIn Performance