Offer
AI workflow automation for businesses
Custom systems that reduce manual work, automate repetitive ops, and turn messy inputs into clean outputs.
Open to the right projects
Haryshwa Ganesh
AI Engineer & Product Builder
I build AI systems that solve real problems — from research papers to production products.
Chennai, India
About
Who I am
A quick snapshot of what I build, study, and care about.
Services
What I can build for you
Offer
Custom RAG systems
Retrieval systems built around your files, databases, and workflows instead of generic chatbot demos.
Offer
Full-stack AI applications
End-to-end products with the backend, interface, and UX needed to make AI useful in real business settings.
Projects
Things I've built
2026
BuiltFreqLoRA-SAM — Polyp Segmentation
Novel dual-pathway LoRA adapter for MobileSAM. Each attention layer gets two parallel low-rank branches — one spatial, one frequency-domain (FFT) — with a learned gate balancing them. Dice 0.927 on Kvasir-SEG with 255K trainable params. Generalises to 3 unseen colonoscopy datasets without retraining. Full-stack app: Next.js point-click UI + FastAPI backend with TTA.
PyTorchMobileSAMLoRAMedical AIComputer VisionFastAPINext.js
2026
Currently BuildingVoyder — AI Vendor Reconciliation
Turns a 2-hour manual accounts payable process into 2 minutes. Parses vendor statements from any format (PDF, Excel, scanned images) using AI, runs a 4-pass invoice matching engine, and outputs reports in the client's existing Excel format. Live client in freight forwarding.
FastAPINext.jsClaude APIPostgreSQLPython
2026
BuiltDataRAG — Enterprise RAG System
Production RAG backend built end-to-end. Hybrid retrieval: BM25 + Qdrant vector search merged and reranked by a cross-encoder. Provider-agnostic — embedder, LLM, and vector store swap via config. Streaming SSE responses, conversation history compression, PDF vision parsing, live financial chart generation, JWT auth with role-based access.
FastAPIQdrantBM25Cross-EncoderPostgreSQLSSEPython
2026
BuiltXCA Vessel Classification & Segmentation
Two-model pipeline for coronary angiography analysis on the ARCADE dataset. YOLOv8n-seg identifies 25 anatomical vessel classes (SYNTAX scoring system). UNet++ with EfficientNet-B3 handles dense vessel segmentation. Full-stack diagnostic UI with annotated overlays and clinical descriptions per detected segment. Ongoing research.
YOLOv8UNet++Medical AIComputer VisionFastAPIPython
Experience
Where I've worked
June 2025 — Present
Levarus Solutions Pvt Ltd
Software Engineer Intern
Chennai, India
- Building an internal RAG system end-to-end — document ingestion, chunking strategies, vector retrieval, reranking, and LLM response synthesis over company files and databases
- Working with Oracle DB in banking workflows — SQL/PL-SQL, indexing, performance monitoring, and transactional data reliability at production scale
- Optimizing query performance and data access patterns to support AI retrieval pipelines in a high-reliability banking environment
2024 — Present
Freelance
AI & Full-Stack Developer
Remote
- Building a full chess education platform for a client — course creation, video delivery, payment integration, and student progress tracking
- Built custom web applications and websites for small businesses, handling everything from design to deployment
- Implemented digital marketing and SEO for client sites, driving measurable traffic growth
Research
Research & Exploration
submitted
FreqLoRA-SAM: A Visual Prompt-Driven Interactive Segmentation Model for Detecting Colorectal Cancer
SRM Institute of Science and Technology · NIT Rourkela · Jadavpur University · 2025 — 2026
First author. Proposed FreqLoRA-SAM, a parameter-efficient adaptation of MobileSAM for polyp segmentation in colonoscopy images. Introduced a dual-domain low-rank adapter combining spatial and frequency-aware LoRA branches with a learnable per-layer sigmoid gate — achieving Dice 0.927 on Kvasir-SEG with only 255K trainable parameters, outperforming methods trained on nearly twice the data. Cross-dataset evaluation on CVC-ClinicDB, CVC-ColonDB, and ETIS-LaribPolypDB demonstrates generalisation without retraining.
submitted
Spectrally Motivated Dynamic Inference: An Empirical Study of Frequency Truncation in Early-Exit Neural Networks
SRM Institute of Science and Technology · 2025 — 2026
Second author. Empirical study of inference-time low-pass filtering on CNN classification accuracy and its application to early-exit architectures. Demonstrated that moderate frequency truncation preserves discriminative capacity while random coefficient retention causes catastrophic degradation — motivating a spectrally-informed early-exit framework using teacher-student distillation and temperature scaling. Achieves 1.60× MAC reduction with under 3% accuracy drop on CIFAR-100.
Writing
Long-form technical thinking.
Deep-dives into AI systems, research, and engineering decisions. No summaries — the full story.
Contact
Let's build something.
Looking for software engineering, AI/ML roles, and meaningful collaborations.
If you want to talk about a role, a product idea, or an AI system for your business, send the details here and I'll get back to you.
© 2026 Haryshwa Ganesh. All rights reserved.