Krimler

Math · Computer Science · AI/ML · Security

Subscribe to the blog feed (RSS) · Reviews feed

Goltzius's painting of Mercury, the messenger

Designing HTTP APIs that AI agents can actually use

ai-agents

http

api-design

ietf

AI agents are a fast-growing kind of API client, and they read an API very differently from a human developer. An IETF draft collects the properties that make an HTTP API easy for an agent to use into one reusable profile.

Jun 9, 2026

Madhava Gaikwad

Counting Fibonacci primes: what a finite computation can and cannot add

number-theory

primes

statistics

reproducibility

Treating the 57 known Fibonacci prime indices as a dataset. A density heuristic, a prediction window for the next one, a battery of tests that finds no hidden pattern, and an exact picture of how primes divide these sequences.

Jun 7, 2026

Madhava Gaikwad

Velazquez's The Spinners, figures weaving thread into fabric

Reading MRC and SRv6: resilient AI supercomputer networking

paper-review

networking

ml-systems

A review of Araujo et al. (arXiv 2605.04333). A bold, candid paper on networking 100,000-GPU clusters, with one clear opportunity: symmetry in the evidence.

Jun 7, 2026

Madhava Gaikwad

Bellotto's view of the Grand Canal in Venice

Reading AccelNet: SmartNICs in the public cloud

paper-review

networking

systems

A review of Azure Accelerated Networking (Firestone et al., NSDI ’18). A landmark paper with production evidence at real scale, and a few places where the evaluation could close the loop.

Jun 7, 2026

Madhava Gaikwad

Lorenzetti's Allegory of Good Government fresco

AEGIS: typed evidence and obligations for AI-native enterprise architecture

ai-governance

enterprise-architecture

ontology

security

Why static architecture diagrams cannot govern AI systems, and a framework that makes evidence, obligations, and runtime signals first-class objects with a regulatory crosswalk.

Apr 9, 2026

Madhava Gaikwad

Domenico Remps Cabinet of Curiosities with many compartments

Did you check the right pocket? Routing retrieval across an agent’s memory

llm-agents

retrieval

memory

ml-systems

Memory-augmented agents keep several specialized stores and usually query all of them for every request. Treating store selection as a cost-sensitive routing problem cuts tokens and improves answers.

Mar 19, 2026

Madhava Gaikwad

Arcimboldo's The Librarian, a figure composed of books

Finite-size guarantees for dense associative memory

machine-learning

associative-memory

learning-theory

Modern Hopfield networks store far more patterns than the classic kind, but the theory only described the infinite limit. This work proves how fast retrieval converges at finite size, how much corruption it tolerates, and why it always settles.

Jan 20, 2026

Madhava Gaikwad

A common language for benchmarking LLM serving: four IETF drafts

llm-systems

benchmarking

ietf

distributed-systems

Throughput, latency, and tokens-per-second mean different things in different papers and products, so the numbers do not compare. A set of IETF drafts fixes the vocabulary, the method, the measurement boundaries, and the workloads.

Jan 20, 2026

Madhava Gaikwad

Ernst Haeckel's plate of sea anemones from Kunstformen der Natur

Twenty-five representative tasks: the IETF workload profiles draft

llm-systems

benchmarking

ietf

The fourth of four IETF drafts on benchmarking LLM serving. It standardizes the workloads themselves, so systems are compared on the same representative tasks, each described along five dimensions.

Jan 20, 2026

Madhava Gaikwad

Declaring the measurement boundary: the IETF profiles draft

llm-systems

benchmarking

ietf

The third of four IETF drafts on benchmarking LLM serving. A profile binds the terms and the method to a concrete role in the system, so a number says exactly what it includes.

Jan 20, 2026

Madhava Gaikwad

Antonello da Messina's Saint Jerome in His Study

A shared vocabulary for LLM serving (IETF terminology draft)

llm-systems

benchmarking

ietf

The first of four IETF drafts on benchmarking LLM serving. It fixes the words: precise definitions for the metrics that everyone uses loosely, with no methodology and no thresholds.

Jan 20, 2026

Madhava Gaikwad

Joseph Wright's An Experiment on a Bird in the Air Pump

How to run the tests: the IETF LLM serving methodology draft

llm-systems

benchmarking

ietf

The second of four IETF drafts on benchmarking LLM serving. It turns the vocabulary into a procedure: where to measure, what traffic to send, and which ten tests to run.

Jan 20, 2026

Madhava Gaikwad

AlignDP: locking knowledge transfer at the data interface

privacy

differential-privacy

llm

security

Watermarking and monitoring react after a model has already leaked. AlignDP blocks knowledge transfer at the input, with strong hiding for rare fields and unbiased frequency estimates for common ones.

Dec 21, 2025

Madhava Gaikwad

Enforcing privacy at the edge for local language models: AVEC

privacy

differential-privacy

llm

edge

A local model that hands hard queries to a bigger remote model can leak private data in the handoff. AVEC proposes adaptive per-query privacy budgets with on-device verification, and proves what such a scheme can and cannot guarantee.

Sep 13, 2025

Madhava Gaikwad

Caravaggio's Narcissus gazing at his identical reflection in the water

When are two RLHF objectives the same? A canonical form for preference losses

ai-alignment

rlhf

preference-optimization

The preference optimization literature is full of objectives presented as distinct improvements. Opal canonicalizes them and shows many are the same objective in disguise, while a few are provably different.

Sep 11, 2025

Madhava Gaikwad

Why the alignment gap always wins, unless you know where to look

ai-alignment

rlhf

learning-theory

When human feedback is subtly wrong on rare cases, learning the true objective takes exponentially many samples. A calibration oracle that flags the unreliable cases collapses that cost to polynomial.

Sep 5, 2025

Madhava Gaikwad

Turner's Rain, Steam and Speed showing a railway in motion

Scoring a datacenter fleet by the risk of what happens next

distributed-systems

reliability

sre

Alerts tell you what already broke. ANSC scores how close a hyperscale network is to running out of capacity, ranked by the probability of an imminent shortfall rather than by current impact.

Aug 22, 2025

Madhava Gaikwad

Vermeer's The Geographer holding measuring instruments

Treating alignment as something you can measure and reduce: NPO

ai-alignment

rlhf

ml-systems

Most pipelines align a model once and ship it. NPO treats alignment as an ongoing quantity, adds a way to check that the monitoring itself is honest, and proves the loop converges under noisy feedback.

Jul 28, 2025

Madhava Gaikwad

Categories