VEERSANGHAVI
Half data scientist. Half AI engineer.
Veer Sanghavi · photographed between two comebacks.
Before the code, there was a court.Before the models, there was a brother.
My brother put a basketball in my hand and a dataset on my screen. I never stopped dribbling either one.
— FOR NISHANT
NISHANT · VEER · KASHMIR · ~2010
The first person who showed me something was cool before the world decided it was.
Arrived as an 18-year-old with a basketball. Left with three papers, two internships, and a builder's instinct.
An 18-year-old from Pune with a basketball, a laptop, and no idea what he was doing yet.
A builder. Three publications in. Two internships shipped. Ready for the US.
VIT gave me two educations.
One in code. One in people.
The technical foundation.
Machine Learning · NLP · Computer Vision · Deep Learning · GenAI · Statistics
Clinical NLP · Synthetic Media Detection · Multimodal Retrieval · Computer Vision
Python · SQL · scikit-learn · XGBoost · PyTorch · TensorFlow · OpenCV · LangChain · LangGraph · FAISS · Transformers (Hugging Face) · NLTK · spaCy
· 2 internships · 1 RAG chatbot before the hype
The human foundation.
V. Sanghavi
Total: Roommates from Agra, lab partners from Kerala, teammates from the Northeast.
Four years. Four milestones.
VIT · Year One
Computer Science, wide-eyed. First loss function that actually trained.
Tredence · First Shot
Built a clinical NLP classifier.
Three Papers Published
Clinical text classification, Drowsiness detection, Facial clustering — three in a year.
Tech Mahindra · Breakthrough
Deepfake detection. Real images vs. AI-generated.
The threewho madethe same leap.
Roots planted.
The Liverpool crest wasn't planned. Neither was any of this.
The Scoreboard
SEASON LINE · 2022–2025
Tredence
- WHERE
- Remote · India
- WHEN
- Dec '22 → Apr '23
- DOMAIN
- Clinical NLP
- STACK
- Python · scikit-learn · SFS · NLTK · SpaCy
classification accuracy. First paper published.
Tech Mahindra
- WHERE
- Pune · On-Site
- WHEN
- Jan '24 → Jul '24
- DOMAIN
- GenAI · CV
- STACK
- StyleGAN2 · XGBoost · ANN
synthetic face detection. Bravo Award.
New York Life
- WHERE
- New York, NY
- WHEN
- May '25 → Aug '25
- DOMAIN
- Multi-modal RAG
- STACK
- ColQwen2 · Amazon Titan V2 · OpenSearch
recall. −800ms latency. No precision drop.
Total Output 2 Papers · 1 Bravo Award · 1 Retriever in Prod
— V. SANGHAVI, BOX SCORE
Each one earned its own page. Turn the page for the long reads.
Teaching a retriever to see and read at the same time.
The TL;DR:Insurance PDFs are messy. Standard AI fails because it only reads text, ignoring layout. Vision AI fails because it doesn't deeply understand words.
The Solution ↓
The Messy Reality of PDFs
Text-Only Retrieval
Standard Embeddings
Destroys spatial semantics like multi-column layouts, tables, and stamps.
Vision-Only Retrieval
Image Embeddings
Fails to understand the dense semantic meaning and domain-specific jargon.
“No single modality can fully understand these documents.”
The Dual-Encoder System
Instead of choosing one approach, we built a parallel dual-encoder system. Both encoders process the document simultaneously, and then their final similarity scores are fused together.
The Interpolation Fusion
The interpolation weight (α) is tuned via grid search per document type.
The Score-Drop Gate
Instead of a fixed Top-K (which passes noise to the LLM), we detect the largest drop between ranked pages and dynamically cut the rest. Fewer pages, faster latency.
The Complementary Fit
Even when both models had similar accuracy, they retrieved completely different pages. ColQwen caught layout, Titan caught semantics.
The hybrid
opens the gap.
| Model | Recall@20 | Recall@30 | Latency |
|---|---|---|---|
ColQwen Vision Only | 77% | 81% | — |
Titan V2 Text Only | 79% | 82% | Baseline |
→ Hybrid α = 0.5 · Interpolated | 82%+3 | 88%+6 | -800ms-20% |
Which one is
real?
Two faces. One is a real person; one was generated by StyleGAN2. I had to look up my own dataset to be sure. That's the problem worth solving — and the one I trained a model to answer in under a second.
THE REVEAL
Both look real. Only one IS real.
Direct image analysis can't catch sophisticated GANs. So we built a forensic detector that doesn't look AT the face — it measures how the face behaves inside the generator's own latent space.
01
The generator gives itself away.
StyleGAN2 can do the reverse of what it's famous for: take any image, invert it back to a latent vector, then regenerate. Fakes come back almost identical— they live inside the generator's native world. Real photos don't come back the same. They drift. They distort. And that drift is measurable.
If a generator can reconstruct you cleanly, you're one of its own.If it can't, you're real.
Nearly identical. Low Euclidean distance, high cosine similarity — the tell for a fake.
Visibly different.The generator distorts what it wasn't trained to produce — and that's the signature.
We measured seven image similarity metrics. Two of them mattered the most.
This is the pipeline that ran it — 10,000 paired images, a short list of features that mattered, and a result that worked cleanly on the problem we defined.
10-fold cross-validation on the 10,000-image paired set. The signal was consistent, not a fluke of one split.
One generator.
One distribution.
The system learned to spot StyleGAN2. Future work: parallel pipelines for BigGAN, CycleGAN, SAGAN.
Guidance: Nikhil Malhotra & Amit Dubey. Makers Lab, Tech Mahindra.
Co-authored with Rashi Solanki.
That's the Tech Mahindra case. One more step back — Tredence · 2022 · the first dataset, the first team, the first time any of this felt real.
Clinical text,
classified.
My first internship. Medical transcripts in, medical specialties out. Classical NLP — tokenize, tag, lemmatize, vectorize — and a feature selection routine that earned its keep.
Five classifiers tested across binary and multiclass settings. All landed in a similar range — and the ceiling dropped fast as classes were added.
The overlap lived in the data.
Cosine and Jaccard similarity showed many transcripts were near-identical across specialties. Doctors write about surgery and history regardless of department.
Data quality beats model choice.
Five different classifiers, one stubborn ceiling. The limit wasn't the model — it was how cleanly the labels separated in the text.
Three internships. One arc.
The shelf.
Five papers between Vellore and the lab. Three published, two whitepapers. Different problems — clinical text, real faces, fake faces, tired faces, university operations — one collaborator across all of them.
2022 – 2024
IEEE · IJRITCC
VIT · TECH M.
CO-AUTHOR
A Retrieval-Augmented Generation (RAG) Chatbot personalized for VIT Vellore Operations
INTERNAL
Detecting Synthetic Human Faces Generated by StyleGAN2
TECH MAHINDRA
Enhancing Driver Safety: Real-Time Drowsiness Detection through Eye Aspect Ratio and CNN-Based Eye State Analysis
CONFERENCE
Clinical Text Classification Using Sequential Forward Selection
IEEE
Data Sharing based on Facial Recognition Clusters
JOURNAL
From Vellore to Tempe.
August 2024. The first time on the other side of the planet. ASU. The shift from doing data science to deciding where to point it.
ASU. The friends in this photo.
A dorm, and a stack of textbooks I didn't always open.
I came in good at the technical side. The other half is what I picked up here.
A/B TESTING · EXPERIMENTATION · STATS & PROBABILITY · ANALYTICAL REASONING · BUSINESS/PRODUCT SENSE
THE GROUP · SEDONA · 2025
This is just where I am right now.
New York. A new chair.
AI Engineer at the American Arbitration Association.
Pune → Vellore → Tempe → New York.
AI ENGINEERING · APPLIED ML · LLM SYSTEM · PRODUCTION · LEARNING ON THE JOB
NEW YORK · WTC · 2025
The skyline from the other side.
Dream Seasons.
THE PORTFOLIO IS NAMED AFTER THEM
Down 3–1 in the series. Iguodala to Curry, back to Iguodala, up for the layup... OH! Blocked by James!
Two weeks after coming back from 3–0 down to Barcelona in the semi final. Corner taken quickly... Origi!
You’ll Never Walk Alone.
Sung at Anfield before every match · the line that holds the whole thing together