LambdaMatters

Our Thesis

The interesting problems live between disciplines.

Mathematics gives us language and structure. Functional programming makes that structure executable. Systems research confronts those ideas with state, distribution, and failure. Machine learning brings empirical search, probabilistic models, and agent workflows. LambdaMatters exists at their intersections.

Σ

Mathematics & foundations

Discrete mathematics, probability, linear algebra, geometry, category-inspired thinking, and computational exploration.

∂

AI & machine learning

Reliable AI systems, generative models, symbolic discovery, and empirical evaluation.

⇄

Systems & protocols

Distributed state, offline-first systems, cryptographic identity, verifiable architecture, and enforceable design intent.

λ

Functional & systems programming

Haskell, Rust, expressive type systems, memory safety, and compositional design.

Research Reports

Report 01 October 2026 Σ Mathematics × ∂ Machine Learning

Newton Reborn: Discovering the Inverse-Square Law with Generative AI

Raghu Ugare

Black-box neural networks fit training curves, not physical laws. While a standard multilayer perceptron reaches R² = 0.940 in-distribution, its accuracy collapses under extrapolation (R² = 0.626 at 1000× range). By coupling an information-theoretic VAE bottleneck (β = 0.4) with symbolic regression via risk-seeking policy gradients, this pipeline rediscovers Newton's exact inverse-square law (F = G·m₁m₂/r²) from 20,000 noisy observations without prior functional assumptions, retaining R² = 0.998 across 1000× extrapolation.

R² = 1.000 Exact Law Recovery

R² = 0.998 1000× Extrapolation

-11.4% VAE Denoising MSE

Paper & Repo (Release Coming Soon)

Cite as BibTeX

@techreport{ugare2026newtonreborn,
  title       = {Newton Reborn: Discovering the Inverse-Square Law with Generative AI},
  author      = {Ugare, Raghu},
  institution = {LambdaMatters},
  number      = {LM-TR-2026-01},
  year        = {2026},
  month       = {October},
  url         = {https://lambdamatters.com}
}

Report 02 September 2026 ⇄ Systems × ∂ AI

Compiled Intent: Enforcing Architectural Invariants for LLM Coding Agents

Vijay Anant

Wikis don't compile. When teams dump entire specification documents into an agent's prompt, models exhibit the Ammunition Effect, using background prose and historical trade-offs to rationalize architectural shortcuts. The Compitent engine indexes architecture specifications into normalized vector and AST databases, retrieving only concise, normative constraint chunks (~750 tokens). Context payload drops by 94.6% while rule enforcement improves by 12.8% to 17.9% across local models.

-94.6% Context Payload

+17.9% Enforcement Gain

~750 Token Footprint

Download Report (PDF) Engine (Rust) ↗ ArchEval Benchmark ↗ Author's Essay ↗

Cite as BibTeX

@techreport{anant2026compiledintent,
  title       = {Compiled Intent: Enforcing Architectural Invariants for LLM Coding Agents},
  author      = {Anant, Vijay},
  institution = {LambdaMatters},
  number      = {LM-TR-2026-02},
  year        = {2026},
  month       = {September},
  url         = {https://lambdamatters.com}
}

Systems & Protocols

Compitent

⇄ Systems × λ Functional

A headless architectural coprocessor and Model Context Protocol (MCP) server that compiles design intent into synchronized Apache Arrow and LanceDB indices.

IDE Integration: Serves targeted constraint chunks to editor agents via the Model Context Protocol (MCP stdio).
Pre-commit & CI Gates: Intercepts unified diffs in git hooks and pipelines to detect boundary violations before merge.
Local & Offline Execution: Operates entirely on-device using LanceDB and Tantivy with zero telemetry or remote calls.

Source (Rust) → Benchmarks →

Akshara

Σ Cryptography × ⇄ Systems

An offline-first, encrypted data architecture for collaborative state without trusted central servers.

Content-Addressed Merkle-DAGs: DAG-CBOR state graphs with causal reconciliation independent of wall-clock time.
Hierarchical Deterministic Identity: Master seed derivation providing functional branch isolation and key revocation walking.
Zero-Knowledge Relays: Content-blind synchronization using XChaCha20-Poly1305 authenticated encryption with forward secrecy.

Source (Rust) → Architecture Series →

AutoCombo

Σ Mathematics × ⇄ Systems

Designed jointly by Raghu Ugare and Vijay Anant. An algorithm mapping product catalogs to prime numbers, collapsing combo selection and multiset containment into single integer divisibility operations.

Prime Factorization Mapping: Assigns unique primes to products, turning shopping carts and combo bundles into unique composite numbers.
Constant-Time Containment Check: Sub-multiset matching reduces to single modulo division (S ≡ 0 mod C).

In the Lab

Semantic Map-Reduce: Auditing Multi-Domain Code Changes

Report 03 (In Development) ⇄ Systems × ∂ AI

When a pull request touches multiple subsystems at once, standard vector retrieval breaks down. Blending unrelated code changes into a single query dilutes the signal, causing the model to miss relevant constraints. We are designing a map-reduce pipeline that breaks complex diffs into cohesive domain fragments, audits each against its local invariants, and aggregates the findings.

Founders

Vijay Anant

vijayanant.com ↗

Raghu Ugare

biscuitsintea.com ↗

2010s: Colleagues, then collaborators exploring programming languages, discrete mathematics, and software design.
2018: Presented (Why?) Should You Know Category Theory? at Functional Conf.
2019: Presented Have You GADT? GADTs for Eliminating Runtime Checks at Functional Conf.
Today: Active research reports, distributed protocols (Akshara), and symbolic physics (Newton Reborn).

Alongside our open research and systems, we work with select teams on hard problems in systems architecture, distributed state, and applied machine learning.

For advisory and technical inquiries: hello@lambdamatters.com