Siddharth N R — teaching small models to think big, distilling frontier intelligence into compact models, building training runs that survive session resets, making language models verifiable and shipping open weights for everyday laptops.Teachingsmallmodelstothinkbig.
AI / MLOps Engineer, founder of Pluto AI Labs, open-source model builder and LLM verification researcher.AI / MLOps Engineer.
Tamil Nadu, IndiaSiddharthFXBuilding Pluto AI Labs
Intelligence isn’t about scale. It’s about precision.
I train, compress and ship open-source models for the hardware people already own — from 68-hour fine-tuning runs on two T4 GPUs to 400 MB coding models that run at 75 tokens/sec on an 8 GB MacBook.
I’m an AI / MLOps engineer and the founder of Pluto AI Labs, an independent research lab shipping open-source models with 10,000+ cumulative downloads on Hugging Face.
I contribute to 8+ open-source AI organizations, including Apple’s MLX, where my released work includes quantized MLX-format models.
I specialize in distributed MLOps, QLoRA fine-tuning and fault-tolerant training on constrained hardware — multi-GPU DDP pipelines, 4-bit NF4 quantization and checkpoints that survive session resets.
Two publications so far, and a growing research focus on LLM verification systems — currently as an AI/ML research contributor at Crebiliti, UK.
I fine-tune and distill small models until they compete with much larger ones — a 3B vision-language model that matches Qwen2.5-VL-3B-Instruct on ChartQA, and a 0.5B coder that outscores DeepSeek-Coder-1.3B on strict EvalPlus.
0.5Bbeats1.3B
Atlas-Coder-2DeepSeek-Coder
strict EvalPlus · 36.6% HumanEval+
02MLOps
Fault-tolerant training pipelines
Distributed DDP + QLoRA pipelines built for ephemeral, constrained compute: 68-hour runs on two T4 GPUs, 4-bit NF4 quantization, and a 3-layer cross-session checkpoint recovery system that fixed a Hugging Face Trainer failure mode.
2× T4 · DDP + QLoRA68 h
0 h68 h
checkpointsession reset · resumed
03Research
Verification you can trust
At Crebiliti I combine deterministic attribute engines with model-based inference — lifting LLM verification accuracy from 78% to 98% on controlled benchmarks, backed by NLI, embedding and LLM-judge experiments.
78%98%
07898
verification accuracy · controlled benchmarks
04Delivery
Shipped, not just prototyped
17 models and 5 datasets on Hugging Face with 10,000+ downloads, packaged for GGUF and Apple MLX — plus a PyPI CLI with CI/CD-ready exit codes for GitHub Actions and Jenkins.
Conduct AI/ML research on evidence positioning for LLM systems, combining deterministic attribute engines with model-based inference to raise verification accuracy from 78% to 98% on controlled benchmarks.
Design and execute controlled experiments across claim–evidence relation categories (NLI, embeddings, LLM judges), producing an attribute-level failure matrix that guides the research roadmap.
Build a validated synthetic-data pipeline for AFU classifier training, combining local LLM generation, dual-NLI validation and leakage-tested dataset splits with full provenance.
NLI
Embeddings
LLM-as-Judge
Synthetic Data
Python
Pluto AI Labs
CurrentIndia · Remote
FounderJan 2026 – Present10 mos
Founded an independent AI research lab; shipped open-source edge models generating 10,000+ cumulative Hugging Face downloads and became an active contributor to 8+ open-source AI organizations, including Apple’s MLX, where released work includes quantized MLX-format models.
Architected a 68-hour 2× T4 DDP QLoRA pipeline for Apollo-VL-Edge-3B, a 3B vision-language model matching Qwen2.5-VL-3B-Instruct on ChartQA (78.6%) and AI2D (78.0%) while outperforming larger 4B+ models — 300+ downloads in 48 hours.
Curated and released Athena-VL-250K, a 250,000-row vision-language corpus powering Athena 1.0-4B, engineered to deliver 9B-class document and chart intelligence on consumer hardware.
Fine-tuned Pluto-Genesis-0.6B via QLoRA to a 44.02% benchmark average (#4 globally among sub-1B models, 2,000+ downloads in 48 hours), solving a Hugging Face Trainer failure mode with a 3-layer cross-session checkpoint recovery system.
Fine-tuned Atlas-Coder-2-0.5B past Qwen2.5-Coder-0.5B and DeepSeek-Coder-1.3B on strict EvalPlus benchmarks (36.6% HumanEval+); packaged as a 400 MB GGUF running at 75 tokens/sec on 8 GB MacBooks.
PyTorch
Transformers
QLoRA
DDP
MLX
GGUF
CT
Codec Technologies
CompletedIndia · Remote
InternshipJan 2026 – Mar 20263 mos
Designed and implemented ERC-1155 multi-token smart contracts on Ethereum with a distributed New Zealand and India team, applying secure infrastructure design that directly informed the peer-reviewed Sentient Audit research.
Developed Solidity smart contracts for Ethereum applications, building deep familiarity with code-level vulnerability patterns.
Solidity
Ethereum
ERC-1155
Contract security
MB
MegaBliss Worldwide
CompletedNew Zealand · Remote
InternshipAug 2025 – Nov 20254 mos
Built real-estate NFT infrastructure on Ethereum with AI-assisted contract logic and automated compliance for an international client, delivering ERC-1155 multi-token smart contracts with a distributed New Zealand and India team.
Ethereum
Solidity
NFTs
AI-assisted logic
IG
Igress Solutions LLP
CompletedBangalore, India · Remote
InternshipJun 20241 mo
Developed LLM-integrated machine learning pipelines for data analysis and automated insight generation, and built Power BI dashboards for client-facing reporting.
Released through Pluto AI Labs on Hugging Face — each engineered to punch above its parameter count and to run where people actually are. Download counts are live.
Vision-language model · 2× T4 DDP + QLoRA · 68-hour run
A 3B vision-language model that matches Qwen2.5-VL-3B-Instruct on ChartQA and AI2D while outperforming larger 4B+ models. Also released as GGUF and 4/6/8-bit + FP16 MLX builds for Apple silicon.
QLoRA fine-tuned to a 44.02% benchmark average. Solved a Hugging Face Trainer failure mode with a 3-layer cross-session checkpoint recovery system — written up as a paper on Zenodo.
Fine-tuned past Qwen2.5-Coder-0.5B and DeepSeek-Coder-1.3B on strict EvalPlus benchmarks, then packaged as a 400 MB GGUF that runs at 75 tokens/sec on an 8 GB MacBook.
A curated 250,000-row vision-language corpus (Athena-VL-250K) powering Athena 1.0-4B — engineered to deliver 9B-class document and chart intelligence on consumer hardware.
“git diff for LLM behavior” · PyPI CLI · Jun – Jul 2026
Behavioral regression testing for language models: a backend-agnostic client (Ollama, OpenAI, Groq), regex-based scorers across behavioral dimensions, and CI/CD-ready exit codes for GitHub Actions and Jenkins.
Real-time, on-chain audits of verified Solidity contracts: Chainlink Functions fetch source from Etherscan, Gemini scores vulnerabilities, and verified results are minted as NFT badges for transparency.
An end-to-end macOS and web platform built solo: authentication flows, on-chain functionality and production-ready deployment pipelines across the full stack.
A browser-only crypto portfolio tracker with live CoinGecko market data and P&L analytics; zero accounts and local-only persistence that auto-expires after 48 hours.
An OCR-to-insight pipeline that reads prescriptions via Google Cloud Vision and uses Gemini to return plain-language clinical guidance, deployed end-to-end on Google Cloud Run.
If you’ve read this far, we probably care about the same things: models that are small, honest and fast — running on the machines people already own. Let’s build the next one together.
Let’s build something precise.
Research collaborations, ML engineering roles, or a hard problem in efficient AI — my inbox is open.
