Research & Innovation Initiative · Since 2025

Abhilash AI Research Lab

The research and innovation initiative of Abhilash Construction Company — studying efficient, accessible, and intelligent AI systems.

What We Study

Efficient, accessible intelligence.

We are studying how much capability can be achieved with small, efficient models under strict parameter and compute constraints — and how to run useful AI on local and edge hardware. This is early-stage research; we report only what we have actually measured.

Research Focus

Research Areas

Directions we are actively studying. Early-stage work is labelled Experimental.

Efficient Intelligence

Achieving strong capabilities with smaller, more efficient models — more intelligence per parameter and per FLOP.

Small Language Models

Highly parameter-efficient language models. A long-term interest is exploring models around the ~10M-parameter range to understand how much intelligence is possible at very small sizes.

Experimental

Edge AI

Models designed to run efficiently on local and edge hardware, without cloud dependence.

Efficient Training

Parameter-efficient fine-tuning (LoRA, QLoRA), quantization, distillation, and efficient architectures for cheaper training.

Native / Indian AI

Early exploration of Indian languages, local deployment, and AI systems optimized for Indian use cases.

Exploratory

What We're Studying

Research Projects

Early-stage research programs. We report only results we have actually measured.

Native Indian LLM

Saffron-v1 — our own efficient ~100M model with a custom byte-level tokenizer and curriculum data: an English foundation built toward Indian-language coverage. A first Preliminary checkpoint is trained and published, plus an Experimental instruction-tuned chat variant.

Own ArchitectureCustom BPE~100MPreliminary

Read more → GitHub → Hugging Face → Chat (Experimental) →

Small Language Models for the Edge

Compact, efficient SLMs designed to run on edge devices with limited compute and power — bringing capable AI directly on-device.

EfficiencyEdge AIQuantization

Architecture Comparison

Systematic head-to-head evaluation of different LLM architectures on our in-house benchmarks to learn what truly works — and why.

TransformersBenchmarkingResearch

Dataset Comparison

Fix the best architecture, train it on different datasets of equal token budget, and cross-evaluate to see which data trains the most capable model.

DatasetsCross-perplexityData-centric

LLM Evaluation Benchmarks

Building rigorous, reproducible benchmarks to measure LLM capability, safety, and efficiency — the yardsticks the field needs.

EvaluationDatasetsReproducibility

Open Data

Open Datasets

Original synthetic instruction datasets we built — programmatically generated with verifiable answers (question → reasoning → answer). Not a benchmark; no third-party LLM used.

Math (5k)

~5,000 arithmetic, algebra, word-problem, and geometry questions, each with a step-by-step worked solution and a computed, verifiable answer.

Worked SolutionsVerifiableSynthetic

Hugging Face →

Reasoning (5k)

~5,000 logic and verbal-reasoning problems — syllogisms, sequences, analogies, coding-decoding — with worked reasoning and answers correct by construction.

LogicDeductionSynthetic

Hugging Face →

English (5k)

~5,000 grammar, vocabulary, and usage exercises — articles, plurals, tenses, agreement, comparatives — from authored rule tables with explained answers.

GrammarVocabularySynthetic

Hugging Face → Generators →

Our open code is live on GitHub — the Architecture Comparison and Dataset Comparison harnesses (shared pure-PyTorch models). More code and models will follow as projects mature; we don't publish results until they're measured.

Collaborate on the research.

We're open to research collaboration, compute sponsorship, and technical discussion.

Get in Touch