👋 Hey, I'm Akash Nath.
I write research, ship products, and try to make deep learning work in places where people say it won't — on a $10 chip, in ancient scripts, inside an MRI scanner, across a 12-lead ECG.
About
I grew up in Assam, Northeast India — a region where good internet arrives late and cutting-edge research arrives even later. That probably explains why I'm obsessed with building AI that works under constraints: on cheap hardware, in ancient languages, with limited data.
I started coding seriously at 17. I graduated from Assam University in 2025, presented a research paper as first author, and deployed an AI model on an ESP-32 microcontroller costing less than a textbook. That project — an AI that classifies fire severity in real time — was presented as an oral and e-poster at the World Summit on Disaster Management, 2025. It was my proof that serious research doesn't need a lab in a rich city.
Since then the work has gone multimodal. My ECG paper fuses raw 12-lead biosignals with a LoRA-adapted BioGPT so the model doesn't just classify a heart rhythm, it writes the clinical note explaining why — accepted at AICTA 2026 at NIT Silchar, with proceedings going to Springer. And I'm building a brain-tumour MRI classifier that tells doctors when it's unsure, because a system that says "double-check this one" is worth more in medicine than one that looks confident. It's under review at IEEE CVMI 2026.
I'm also training a GPT-2 for Sanskrit — because one of the world's oldest written languages deserves a language model, and nobody else in my corner of the world was doing it. When I'm not doing research, I'm shipping products, competing on CodeChef (4-star), and applying for M.Tech AI/ML programs.
Now
A frozen 1D-ResNet encoder reads 12-lead ECG, projects it into token space, and hands it to a LoRA-adapted BioGPT — so the system classifies the rhythm and generates the clinical explanation, with GradCAM++ saliency over the signal. Accepted at AICTA 2026, NIT Silchar.
Most AI just gives you an answer. MEDIAX gives you an answer plus a confidence score — and when it's not sure about a glioma, it says so. That's the feature. Submitted to IEEE CVMI 2026.
Sanskrit is one of humanity's oldest knowledge systems and it has almost no AI tooling. I trained a 97.7M-parameter language model on the AI4Bharat Sangraha corpus with a custom Devanagari BPE tokeniser. Evaluation ongoing.
A CNN-BiLSTM forecaster with SHAP attribution decomposed across seasonal and diurnal axes, trained on CPCB hourly data from four Indian metros. It puts a number on the intuition: 08:00 IST is the peak attribution window for PM₂.₅ and PM₁₀ — empirical backing for morning traffic policy.
Published work
Fuses 12-lead ECG with a language model: a 4.2M-parameter 1D ResNet encoder projects the signal into BioGPT's token space, with LoRA adapters training roughly 0.5% of the LLM's weights. Trained on 18,750 PTB-XL recordings across five diagnostic superclasses. Multi-token fusion lifts accuracy to 72.6% (+5.6 points over the ECG-only baseline) at macro AUC 0.874, generates clinical text at BERTScore-F1 0.989, and GradCAM++ faithfulness scoring confirms the saliency maps track what the model actually uses.
A customised EfficientNetB4 backbone with Squeeze-and-Excitation attention, built to separate visually near-identical Asian turtle species for conservation monitoring. Beats 11 state-of-the-art baselines at 97.25% accuracy; soft-voting with YOLOv8 adds a further 2.41%. Grad-CAM shows the model keys on the morphological regions a herpetologist would actually look at.
A hybrid EfficientNetB3 + BiLSTM that classifies six fire-severity levels from a live camera stream in under 500ms — deployed on ESP-32 hardware. 94.39% accuracy on the MIVIA dataset. Presented as first-author oral and e-poster at the World Summit on Disaster Management, Graphic Era University, 2025.
EfficientNet-B3 with multi-scale feature fusion and Monte Carlo Dropout. Hits 93.75% accuracy and AUC 0.983. The real contribution: a clinical triage layer that drops glioma miss-rate from 22% to 2.8% by trading some recall for safety. Built for clinicians, not benchmarks.
CNN-BiLSTM beating six baselines on the Delhi test set (RMSE 13.83, R² 0.9874), with SHAP GradientExplainer decomposed jointly across season and hour. Evaluated across four climatic zones — including an honest account of where it fails (Chennai R² 0.68, unmodelled sea-breeze meteorology). Code, figures and result tables are public.
97.7M parameters, trained on the AI4Bharat Sangraha corpus. Custom 16K Devanagari BPE tokeniser. Evaluation across perplexity, type-token ratio, and script purity. Composite eval score 7.55/10 at checkpoint 83,000. Draft paper targeting ACL/EMNLP/LREC-COLING.
Selected builds
Seven-phase pipeline: preprocess PTB-XL, train a 1D ResNet ECG encoder, project the embedding into BioGPT's space, fine-tune with LoRA on an 8 GB laptop GPU, then explain it with GradCAM++ and cross-attention maps. Ten hours of training, two fusion variants, a full ablation — and a repo you can actually re-run end to end.
EfficientNetB3 + BiLSTM hybrid that classifies six fire severity levels from an ESP-32 CAM stream in under 500ms. Uses a Gemini-assisted + human-expert labelling pipeline on the MIVIA dataset. First-author at WSDM 2025 in Dehradun. This is the one that got me into research.
Built from scratch in PyTorch — multi-scale EfficientNet-B3 fusion, MC-Dropout, temperature scaling, two-phase training. The clinical triage layer is the real innovation: it cuts glioma miss-rate from 22% to 2.8% by refusing to make a confident call when the model is genuinely unsure. Submitted to IEEE CVMI 2026.
CNN-BiLSTM forecaster over six pollutants and six years of CPCB hourly data from Delhi, Mumbai, Kolkata and Chennai. SHAP attribution is split across season and time of day, which turns a black box into something a policy team can argue with. Component ablation shows the CNN block alone buys a 23.2% RMSE reduction.
97.7M-parameter transformer trained on the AI4Bharat Sangraha corpus with a custom 16,000-token Devanagari BPE tokeniser. Trained iteratively across RTX 2060, T4, and A100 environments with a tuned bf16 + AdamW + NVMe-cached pipeline. One of the few Sanskrit LLMs built from first principles.
Attention-enhanced EfficientNetB4 with Squeeze-and-Excitation blocks, trained on a curated Turtle1 dataset plus images.cv. 97.25% accuracy against 11 benchmarked architectures, and a soft-voting ensemble with YOLOv8 pushes it 2.41% further. Grad-CAM makes the species call auditable. Published in IJEEI, June 2026.
A vision-language pipeline that generates Sanskrit captions from images using CNN-LSTM architectures. Sits at the intersection of the two hardest constraints I keep choosing: multimodal grounding, and a language with almost no paired training data to ground it against.
Image restoration, recolouring, and semantic search in one platform. Clerk for auth, Stripe for subscriptions, MongoDB for persistence, and a credit-based metering system that makes inference cost predictable. This is what "shipping" actually looks like — not just a Colab notebook.
Where I've been
Tools
Highlights
FAQs
Get in touch
Whether it's research, an M.Tech opening, a consulting chat, or just wanting to nerd out about AI — my inbox is open.
I'm based in Assam, India. I reply to emails. I'm also active on LinkedIn and GitHub — the links are all right here. If you're a researcher, recruiter, or founder building something in AI, I'd love to hear about it.
Email me →