I'm an ML engineer who fine-tunes small language models on consumer hardware — and builds the products around them. Most of my work sits at the messy intersection of applied research and things people actually use: clinical AI for frontline health workers, offline-first LLM pipelines, and cross-modal transfer learning.
My paper on cross-modal LoRA degradation was presented at IMPACT 2025 (Springer). A second one, on preventing catastrophic forgetting, is in preparation.
Systems Dev intern @ Cognavi · Uttar Pradesh, India · Portfolio
Cross-Modal Transfer Learning in Domain-Adaptive Video Summarization IMPACT 2025, Springer — presented Dec 6, 2025 · proceedings forthcoming
I found that LoRA fine-tuning on academic papers degrades video summarization by 14–50%, then built a 5-method curriculum framework to fix it.
| System | Video ROUGE-1 | Video ROUGE-2 | CMC |
|---|---|---|---|
| Baseline (zero-shot) | 0.263 | 0.032 | 0.356 |
| IMPACT 2025 naive LoRA | 0.272 | 0.060 | — |
| Full framework (Phase 3) | 0.417 (+58%) | 0.119 (+272%) | 0.531 |
p < 0.005, paired t-test · bootstrap 95% CI · 1,400-line eval harness
Second paper in preparation: Preventing Catastrophic Forgetting in Cross-Modal Summarization.
Clinical AI
| Project | What it does | Highlights |
|---|---|---|
| Sakhi | Offline-first Hindi voice-to-form for India's 1M+ ASHA health workers — home-visit conversations become structured government forms with real-time maternal & child referral decisions. | 87% end-to-end pass on synthetic Hindi audio · dual deploy: health-center (~15s) + on-device Android via Cactus SDK (~5min) · 6-layer anti-hallucination pipeline · Gemma 4 Good Hackathon |
| MedScribe | Speech → concise SOAP notes in physician shorthand, fully offline on 16GB VRAM. Built for the Google MedGemma Impact Challenge 2026. | 16 → 5 min per encounter · 0% hallucinated findings · 90/100 quality · LoRA-tuned on 712 samples for $1.28 |
Generative AI
| Project | What it does | Highlights |
|---|---|---|
| YouTube Transcript Summariser | Domain-adaptive summarization for 90-minute lectures on a single 8GB GPU. The negative result that started the research. | Published — IMPACT 2025 (Springer) · documented a 14–50% cross-modal ROUGE drop, p<0.001 |
| Hybrid Dataset Summariser | The follow-up that fixes it — a 3-phase curriculum (LoRA+, OPLoRA, EWC, CrossCLR) on Mistral-7B over a 4,324-sample cross-modal dataset. | Video ROUGE-1 0.417 (+58%), ROUGE-2 +272% · forgetting held at 14.1% vs 31.4% · all gains p<0.005 · HF model · dataset |
| Bawarchi | Multimodal recipe generation: vision detects ingredients, embeddings + PMI suggest cuisine-appropriate swaps, a fine-tuned LLM writes the recipe. | YOLOv8m 66.5% mAP · substitution engine 85% precision · Llama 3.2 3B tuned on 335K recipes (74.1% accuracy) |
Data Products
| Project | What it does | Highlights |
|---|---|---|
| DPDPA Compliance | Legal-tech automation for India's Digital Personal Data Protection Act, 2023 — 50+ regulatory requirements assessed in minutes instead of days. | 40 hours → 5 minutes (480× the manual baseline) · ~95% PDF extraction accuracy · 9-table relational DB · 15-question assessment engine · targets the May 2027 deadline |
| Tech Layoffs Dashboard | Interactive analytics on global tech-sector layoffs, 2020–2025, with six custom metrics (efficiency, instability, severity, and more). | Live Streamlit · YoY/QoQ trend analysis · company-level deep-dives · Kaggle EDA |
llmclean — zero-dependency Python library for cleaning and normalizing raw LLM output · PyPI
Model Adapters — trained LoRA adapters for MedScribe, Bawarchi, and the Hybrid Summariser · Hugging Face
Datasets — curated training datasets for all projects · Kaggle
Beyond the badges: Transformers · PEFT / LoRA · QLoRA · Unsloth · bitsandbytes · OPLoRA · EWC · Whisper · YOLOv8 · Mistral · Gemma · Llama 3.2
Portfolio · Hugging Face · LinkedIn · Kaggle · tusharbrisingr9802@gmail.com