Selected work.
Seven systems shipped — clinical AI, generative pipelines, data products. Scroll down to pan across.
Open to · ML & product roles
Hi, I'm Tushar.
ML systems engineer
I build ML systems and ship the products around them — on-device LLMs, clinical AI, and data tooling that makes policy actionable.
01 — About
About
A short read on how I work and what I've built.
I build at the intersection of research and shipping: fine-tuning small LLMs, fighting catastrophic forgetting, designing offline-first pipelines, and packaging it all into something a non-engineer can actually use.
Recent work spans clinical documentation (MedScribe), Hindi voice-to-form for ASHA workers (Sakhi), cuisine-aware recipe generation (Bawarchi), and a published method for cross-modal LoRA training (IMPACT 2025).
Currently
Backend Engineer · Cognavi
Publication
Cross-Modal Transfer Learning in Domain-Adaptive Video Summarization
7
ML & data systems shipped
300+
DSA problems solved
Top 7%
Anaverse 2.0 · 1,100+ entrants
model_card.yaml
- arch
- "ML systems engineer + product builder"
- focus
- [on-device LLMs, clinical AI, cross-modal transfer]
- frameworks
- [PyTorch, Transformers, PEFT, Unsloth, Cactus]
- hardware
- [RTX 5070 Ti, T4 (Colab), 8GB consumer GPU, Android NPU]
- biases
- [ships > polishes, smaller > bigger, offline > cloud]
$ ./stats --live
fetching…
02 — Skills
Feature set.
Five shards. Run the script to load them.
tushar@portfolio:~/features$
python reveal_skills.py
Click or press ↵ Enter to run
Languages
- Python
- C++
- SQL
- JavaScript
- Bash
ML & LLMs
- PyTorch
- Transformers
- PEFT / LoRA
- BitsAndBytes
- Unsloth
- Whisper
- YOLOv8
- Quantization (4/8-bit)
Data
- Pandas
- NumPy
- Plotly
- Matplotlib
- Seaborn
- SQLite
Build & Deploy
- FastAPI
- Streamlit
- Gradio
- React
- Vite
- Hugging Face Hub
Workflow
- Git / GitHub
- Jupyter
- Colab
- VS Code
- WandB
Sakhi
ASHA voice-to-form · Hindi
Offline-first voice-to-form for India's 1M+ ASHA health workers. Hindi home-visit conversations become structured government forms with real-time referral decisions for maternal and child health.
Dual deployment: health-center mode (Whisper ASR + Gemma 4 E4B function-calling, ~15s/visit) and field mode (on-device Gemma 4 E2B INT4 via Cactus SDK on Android, ~5min/visit). Six-layer anti-hallucination pipeline with clinical range validation. Fine-tuned on 1,154 synthetic ASHA conversations, then stress-tested on real Hindi voice (4 recordings, 2 speakers).
- 87% end-to-end pass rate (synthetic Hindi audio)
- 14/15 fine-tuned eval · 15/15 base
- Gemma 4 Hackathon — Health, Ollama, Unsloth, Cactus tracks
Python FastAPI Ollama Gemma 4 Whisper React Capacitor Cactus SDK
MedScribe
Clinical SOAP notes · LoRA
AI clinical documentation that produces concise SOAP notes in physician shorthand — not verbose AI-scribe output that needs editing. Built for the Google MedGemma Impact Challenge 2026 (main + novel-task tracks).
MedASR (105M Conformer) for speech, MedGemma 1.5 4B with a 4.2M-parameter LoRA for note generation. Six clinical intelligence tools: billable diagnosis extraction, completeness checks, differential ranking, drug interactions, intake analysis. Fully offline on 16GB VRAM.
- 16 → 5 min per encounter — ~1,100 extra patient visits/year per physician
- 90/100 quality · 0% hallucination · 46% shorter notes
- LoRA fine-tuned on 712 curated samples for $1.28
Python MedGemma LoRA 4-bit quantization Gradio Conformer ASR
Bawarchi
Role-aware recipe generation
Multimodal recipe generation pipeline: vision detects ingredients, embeddings + PMI suggest cuisine-appropriate substitutions, and a fine-tuned LLM writes the recipe. Same ingredients yield tacos (Mexican) or roti wraps (Indian) depending on cultural context.
- YOLOv8m ingredient detection — 66.51% mAP across 124 classes
- PMI + embeddings substitution — 85% precision
- Llama 3.2 3B fine-tuned on 335K recipes — 74.1% accuracy
- 1K curated role-based examples; 43K filtered synthetic (54% accept)
PyTorch Transformers PEFT YOLOv8 Sentence-Transformers Streamlit CUDA
YouTube Transcript Summariser
Cross-modal LoRA · published
Domain-adaptive summarization for 90-minute technical lectures: Mistral-7B + LoRA, 4-bit quantization, streaming chunk processing — fits on a single 8GB GPU.
Published as "Cross-Modal Transfer Learning in Domain-Adaptive Video Summarization" (IMPACT 2025, Springer, Dec 6 2025) — the first work to quantify that LoRA adapters trained on academic papers degrade video summarisation by 14–50% (p<0.001). Hybrid Dataset Summariser is the follow-up that fixes it.
- Published — IMPACT 2025 (Springer), Dec 6 2025
- Cross-modal degradation documented: 14–50% ROUGE drop, p<0.001
- Whisper anti-hallucination ASR preserves ~95% of content
Python PyTorch Transformers PEFT Whisper Streamlit
Hybrid Dataset Summariser
Three-phase curriculum · follow-up to IMPACT
Solves the catastrophic forgetting documented in IMPACT 2025. Three-phase curriculum on Mistral-7B: P1 LoRA+ asymmetric LRs on papers, P2 OPLoRA + EWC mixed modalities, P3 CrossCLR contrastive alignment for video.
Curriculum phasing alone delivers ~80% of the gain over vanilla LoRA. OPLoRA preserves semantic quality (BERTScore +45%). All improvements significant at p<0.005.
- Video ROUGE-1: 0.417 — +58% baseline, vs 0.272 IMPACT naive
- Video ROUGE-2: 0.119 — +272% vs baseline
- Forgetting held at 14.1% vs 31.4% IMPACT
- Best ablation (no EWC): R-1: 0.438, R-2: 0.137
Mistral-7B QLoRA NF4 OPLoRA EWC CrossCLR WandB
DPDPA Compliance
RegTech · India DPDP Act
Legal-tech automation for India's Digital Personal Data Protection Act, 2023. 50+ regulatory requirements processed at 480× the manual baseline (40 hours → 5 minutes). Targets the 1,000+ SMBs facing the May 2027 compliance deadline.
- 9-table relational database — requirements + assessments
- PDF extraction pipeline at ~95% accuracy
- 15-question assessment engine with Excel reports
- Penalty exposure analysis: ₹10L → ₹250Cr
Python Streamlit SQLite Plotly PDF Processing
Tech Layoffs Dashboard
2020–2025 · custom metrics
Interactive analytics dashboard exploring global tech-sector layoffs with six custom metrics: Layoff Efficiency Score, Instability Score, Severity Index, Location Fragility, Industry Survivability, Bounceback Potential.
Quarterly trend analysis with YoY/QoQ tracking, company-level deep-dives, normalized per-company averages for fair cross-scale comparison. Sourced from verified Kaggle layoff trackers (Bloomberg, TechCrunch, NYT).
Python Pandas Plotly Streamlit Jupyter
End of rail.
Get in touch →
04 — Open Source
Open source.
The projects are what I built; these are the pieces I pulled out, hardened, and published — pip install-able, and running in other people's pipelines.
llmclean
v0.3.0Cleans and normalizes raw LLM output — JSON repair from prose, fence & reasoning-trace stripping, repetition trimming, typography. Scoped from a five-model local sweep across llama3.1 / gemma4 / qwen2.5 / deepseek-r1 / mistral.
$
pip install llmclean
from llmclean import enforce_json, strip_reasoning_trace enforce_json('sure: {"ok": True, "items": [1,2,3,]}') # → {"ok": true, "items": [1, 2, 3]} strip_reasoning_trace('<think>hmm…</think>\nParis') # → Paris
hindi-normalize
v0.1.0Turns messy Hindi ASR/OCR into clean text — number-words → digits (tolerant of ASR spelling variants), Devanagari normalization, and medical-term mapping. The speech-to-form cleanup from Sakhi, packaged.
$
pip install hindi-normalize
from hindi_normalize import normalize_transcript normalize_transcript('आपका BP एक सौ दस बटा सत्तर है, वजन अट्ठावन kg') # → आपका BP 110/70 है, वजन 58 kg
Models
-
MedGemma 4B → SOAP notes in physician shorthand · 4.2M-param LoRA · 46% shorter · 0% hallucination
-
Llama 3.2 3B → role- & cuisine-aware recipe generation · fine-tuned on 335K recipes · 74.1% acc
-
Mistral-7B → cross-modal paper→video summarization · QLoRA · 3-phase curriculum · +58% ROUGE-1
Datasets
-
712 curated encounter → SOAP-note training samples
-
1,154 synthetic Hindi ASHA home-visit conversations for voice-to-form fine-tuning
-
1K role-based + 43K filtered synthetic recipe examples
-
Cross-modal summarization data — academic papers + video transcripts
05 — Experience
Experience
Where the work has been put to use — and the research that shaped it.
06 — Contact
Contact
Easiest way to reach me — pick whichever you prefer.