portfolio · linkedin · leetcode · email
Abstract. I'm a final-year computer science student in Bengaluru and an AI/ML engineer intern at Tapplent, where I work on enterprise RAG. Most of what I build ends up in front of someone who can overrule it: a radiologist, a pharmacist, a network operator, an HR admin. So I spend as much time on the evidence behind an answer as on the answer. I also contribute upstream, with five pull requests merged into Next.js, Hugging Face Transformers, Cloudflare's Workers SDK and Apicurio Registry. I've also co-authored two research papers.
Keywords: open source, retrieval-augmented generation, explainable AI, computer vision, serverless systems.
Five merged pull requests in four projects. They're small on purpose, and two got smaller in review: Cloudflare's maintainers had me drop a set of migration warnings, and Hugging Face had me delete my test. Both calls were right.
Each card below links to its pull request.
In review: pytorch/pytorch #188945, assert_close comparisons with Python and NumPy scalars.
AI/ML Engineer Intern, Tapplent (Sep 2026 – present). I'm building a RAG chatbot over 1,000+ HR policy documents for Tapplent's HCM clients. Retrieval combines dense vectors with BM25 and re-ranks with a cross-encoder, which cut retrieval noise and hallucinations by over 40%. Answers cite their source documents, and metadata filters keep each user inside their access level.
Before this: Cloud & DevOps Trainee at Torii Minds (Aug 2025 – Jan 2026), working on AWS and GCP infrastructure and CI/CD.
Two questions shape most of my design decisions. Can the person relying on this check it? And does it still work when conditions are bad?
| Project | What it does | How you can check it |
|---|---|---|
| DeepBreath AI | Classifies chest X-rays as normal, bacterial or viral pneumonia with a five-model ViT ensemble (~94–96% accuracy), then estimates 3D consolidation volume from the single 2D image. | Grad-CAM++ heatmaps are masked to segmented lung fields, so attention can't land on the heart or chest wall. Findings map to five named lobes. |
| CardioXAI | Stratifies cardiovascular risk with a RF, XGBoost, LightGBM and TabPFN ensemble (cross-validated AUC up to 0.96 on UCI Cleveland), under 15 ms on CPU. | A per-patient SHAP waterfall in under 10 ms, model disagreement reported as confidence, and a PDF report a clinician can keep. |
| MedLens AI | Reads handwritten prescriptions with TrOCR, resolves drug names against RxNorm and flags interactions. Gemini runs only when extraction fails schema validation. | Low-confidence results stop the pipeline. A review screen makes a person confirm the drug before anything downstream runs. |
| MFD-AI | Detects manipulated images, video, audio and text in one consensus verdict. | The verdict is built from separate forensic signals (ELA, PRNU sensor noise, FFT artefacts, optical flow, lip-sync), and all 35 analysis stages stream to the operator. |
| Project | The conditions | What happened |
|---|---|---|
| IKYA 2026 (live) | A week of bursty registrations for an intercollegiate fest, with payments and jury scoring on the same platform. | Ran on Cloudflare Workers, D1 and R2: 87,017 requests, 0 errors, 1.1 ms median CPU time, 2,500+ participants from 60+ colleges. |
| GramSetu AI | Rural areas with unreliable internet, and photos people shouldn't have to upload. | YOLOv8-nano and OCR run in the browser through ONNX Runtime Web and Tesseract.js, so images stay on the device. |
| HostelDesk | Students sharing screenshots of their gate passes. | QR codes are signed tokens that rotate every 60 seconds, across a seven-role outpass workflow. |
| RoadSoS | Someone at a crash site, under stress, on a weak signal. | One tap finds nearby hospitals and police. Results are cached for offline use, and it falls back when GPS or the map API fails. |
[1] Niranjan L, Mahalakshmi C K, Mahesh Kumar Jha, Vijay Anand P, Sajithra Varun, Partha Shankar. A Multi-Agent Reinforcement Learning Framework with Graph Neural Networks for Congestion-Aware Routing in Next-Generation Networks. 2026.
Combines CTDE multi-agent RL, a graph attention encoder and SHAP explanations for routing decisions. Against AI-CARF at full load: 36% lower latency (62 ms vs 97 ms) and 46% less packet loss. My part: the congestion index and reward function, the GAT encoder and the SHAP pipeline.
[2] Niranjan L, Ann Mathews, K. Ezhilarasan, Sajithra Varun, Mahalakshmi C K, Partha Shankar. A Deep Learning–Driven Intelligent Crop Prediction Model Integrating CNN–LSTM Architecture for Smart Agriculture. 2026.
A CNN–BiLSTM with temporal attention that recommends crops from soil, climate and satellite data, and names the factors behind each recommendation. My part: the multimodal data pipeline, SMOGN augmentation, the attention layer and hyperparameter tuning.
Also: C++, SQL, LangChain, LangGraph, MongoDB, AWS, Google Cloud.
- Winner, Vibexathon (national hackathon): team lead, MFD-AI
- 2nd place, Nexora (national hackathon): team lead, MedLens AI
- Round 3 qualifier, Flipkart GRiD 8.0, AI Engineering track
- AWS Certified: AI Practitioner (AIF-C01), Solutions Architect Associate (SAA-C03)
- B.E. Computer Science and Engineering, Nagarjuna College of Engineering and Technology, 2023–2027, CGPA 9.00/10
Correspondence: parthashankar21@gmail.com



