Photography
Through
the lens
Personal collection · travels & moments
Building explainable, scalable AI systems, from frontier research to real-world deployment. Specialising in medical image analysis, physics-guided vision, and production LLM systems.
AI Research Architect & Developer
Jun 2026 — Present
AI Engineer / Researcher
Oct 2022 — Present
I'm an applied researcher and engineer with 4+ years of experience, holding two posts at once: AI Research Architect at GentracLabs and AI Engineer / Researcher at TCS Innovation Labs. My work translates novel architectures from research papers into systems that actually ship at scale.
At GentracLabs I build the non-invasive clinical imaging stack: colour science that turns any phone camera into a measuring instrument, and screening models that carry the physics of light and tissue inside the training objective rather than leaving a network to rediscover it. At TCS the work spans production LLM systems, medical computer vision, weakly supervised deep learning, biometrics, and edge AI.
Seven peer-reviewed publications across IEEE, Springer, and Elsevier, with two works currently under patent evaluation.
I care deeply about building AI that is explainable, efficient, and genuinely useful, not just benchmark-optimised.
Concurrent roles: clinical imaging research at GentracLabs, applied AI at TCS Innovation Labs.
Architect and builder of the non-invasive clinical imaging stack, screening disease from an ordinary photograph rather than a blood draw. Three end-to-end systems delivered inside the first months: a physics-based colour-rectification engine, a neonatal jaundice screener, and an ocular anemia screener. The through-line is that the optics live inside the method. Light transport and tissue reflectance shape the objective itself, instead of being left for a network to rediscover from data.
Built NormaEngine, which uses no neural network anywhere, only optics and linear algebra. Ambient subtraction from a flash/no-flash RAW pair, an exposure-invariant root-polynomial CCM into CIE XYZ/Lab, and a per-sensor profile registry so any phone becomes a measuring instrument.
Neonatal jaundice from a whole-body photograph. Melanin and bilirubin scatter light differently, so that difference is written into the training objective, forcing the model to separate skin tone from bilirubin instead of correlating with it.
Haemoglobin estimated from conjunctival pallor, then thresholded against WHO cutoffs by sex and age. Regressing first and classifying second is what holds the model steady across cohorts at 72% and 19% anemia prevalence. SimCLR self-supervised pretraining on unlabelled crops, using pallor-preserving augmentations only, hardens the encoder against capture conditions. Cross-site AUC 0.90, training in India and testing in Italy.
Working across the full AI stack, from publishing novel architectures in top-tier venues to deploying them in production systems used by millions. Key focus areas: generative AI pipelines, medical computer vision, biometric identity systems, NLP, and edge AI. Two works under patent evaluation at TCS (2026).
Architected an enterprise Hybrid RAG pipeline for Passport Seva policy intelligence using GPT-5.4 Pro / DeepSeek-V3, NV-Embed-v2 embeddings, Qdrant vector DB, and Qwen3-Reranker-8B. Integrated Lakera Guard, FSDP2, and vLLM PagedAttention for production-grade GenAI.
Integrated multimodal biometrics (fingerprint, signature, facial) into Passport Seva with UIDAI/Aadhaar real-time auth. Built DigiYatra facial recognition for seamless airport check-in with biometric tokenization at scale.
Segmentation: Lightweight encoder-decoder with boundary-aware loss · 91% mIoU · generalised to 3 domains
NER Pipeline: SpaCy on 120K+ logistics docs · 89% F1 · 180ms → 10ms (18× via quantization + x86 export)
IEEE · Springer · Elsevier
Proposed a novel Confidence-Aware Probabilistic Class Activation Map architecture for multi-disease localization in chest X-rays without annotated labels. Achieved 94% IoBB and 90% Dice score across 13 thoracic diseases on NIH and Stanford CheXpert datasets.
Designed CX-Ultranet with channel shuffling, skip connections, and multiclass cross-entropy loss for simultaneous classification of 13 thoracic disorders. Outperformed SOTA models by up to 15% with 30% less training time.
Modified SRGAN with Residual-in-Residual Dense Blocks (RRDB) for license plate super-resolution integrated with an OCR pipeline. Improved character detection accuracy from 4–6% to 84–90% on severely degraded images.
Early framework combining CAN and ADN for weakly supervised anomaly localization on large-scale CXR benchmarks. Extended into full CAPCAM journal paper.
Original CX-Ultranet architecture for 13-class thoracic disease classification. 88% accuracy with significantly reduced FLOPs; foundation for the extended journal paper.
Applied SIRD epidemiological modelling to COVID-19 data from 37 US states. Demonstrated hard lockdown followed by disciplined soft lockdown significantly curbs R₀.
Probabilistic graph analysis of random vs. planned visitor traversal across Durga Puja festival pandals. Modelled stochastic and deterministic hopping strategies to compare unique visit counts.
Colour rectification for clinical imaging with no neural network anywhere, only optics and linear algebra. A flash/no-flash RAW pair cancels the ambient light, an exposure-invariant root-polynomial CCM maps the now-known flash illuminant into CIE XYZ/Lab, and a per-sensor profile registry pools spectrophotometer-calibrated sessions so any smartphone reports true tissue colour rather than its own interpretation of it. Where a capture cannot support a trustworthy reading, a consistency gate refuses it instead of guessing.
Real-world cross-device error: mean ΔE2000 14.2 → 5.9 across 600 measurements (6 phones × 5 lighting conditions); ΔE2000 < 2 on the controlled path.
Screens newborns for jaundice from a whole-body photograph. The novelty is physical rather than architectural: melanin and bilirubin absorb and scatter light differently, and that optical difference is written into the training objective itself, so the model is forced to separate skin tone from bilirubin instead of quietly correlating with it. A cephalocaudal bilirubin field is fitted per infant, matching the head-to-toe progression clinicians already grade by, with illuminant normalisation and causal disentanglement upstream.
Estimates haemoglobin from the pallor of the eye's conjunctiva, then flags anemia against WHO cutoffs stratified by sex and age. Regressing Hgb and thresholding afterwards, rather than classifying directly, is what keeps it stable across cohorts with 72% and 19% anemia prevalence. SimCLR self-supervised pretraining on unlabelled crops (pallor-preserving augmentations only) hardens the encoder against capture conditions.
Cross-site AUC 0.90, training in India and testing in Italy.
Enterprise-grade policy Q&A system. Contextual semantic chunking, metadata-aware retrieval, Qwen3-Reranker-8B reranking. Integrated Lakera Guard (prompt injection) + vLLM PagedAttention for low-latency production deployment. Significantly reduced hallucinations in long-context policy queries.
Weakly supervised multi-disease localization across 13 thoracic pathologies. 94% IoBB, 90% Dice on NIH + CheXpert. Published in Elsevier BSPC 2024.
Facial recognition for airport check-in. Biometric tokenization with cross-verification against Aadhaar/UIDAI at scale. Reduced manual verification overhead significantly.
SpaCy NER trained on 120K+ logistics documents. 89% F1 across 8 entity types. Quantized for x86 edge: 180ms → 10ms inference. Eliminates cloud API dependency.
Lightweight encoder-decoder with boundary-aware loss. 91% mIoU on facial benchmarks, generalised across 3 additional domains.
10+ years of billing data. Real-time anomaly monitoring with tiered alerting (10L early warning / 50L critical) for enterprise-scale financial deviation detection.
Hover nodes to explore
Personal collection · travels & moments
Custom boundary-aware loss function · 91% mIoU on facial region benchmarks · Generalised across 3 additional segmentation domains beyond the original task
Trained on 120K+ annotated logistics documents · 89% F1 across 8 entity types · Quantized & exported to x86 edge hardware
Interested in collaboration or research opportunities?
anweshkabiraj@icloud.comOpen to research collaborations, AI engineering roles, and applied ML consulting engagements.
📍 Kolkata, India · +91 89670 87246
