Charan Sai Ponnada

I joined Aynstyn Technologies as an AI Engineer intern and converted to full-time Software Engineer on a PPO. I built LabSoft (Next.js, PostgreSQL) end-to-end as the sole developer, shipping the Patient Management and Reporting modules from schema design through deployment, and architected Aynstyn Intel, an internal analytics platform with Bloom's Taxonomy distribution tracking and knowledge-gap heatmaps. On the AI side I improved LLM response reliability across 3+ production workflows with systematic black-box test suites, and do prompt tuning and evaluation on our in-house agent built on Amazon's open-source Strands Agents SDK. I also own development and CI/CD for the React Native mobile app.

I am building a genomic foundation model from scratch on a 2× NVIDIA L40S cluster — a HopField-Mamba hybrid that conditions pretraining directly on 3D chromatin (Hi-C) structure instead of adding it post-hoc. The structural bias lives inside the SSM recurrence at +0.43% parameter overhead, under a matched-compute constraint. Along the way I found a memory-horizon collapse in Mamba's default timestep initialization; fixing it raised median effective memory span ~30× with validation loss unchanged within seed noise, which is being written up on its own. I also have a sole-author paper under review at IEEE InCODE 2026 on semantic consistency as an unsupervised hallucination signal in LLMs.

2nd place among the Top 10 finalists at the IIT Hyderabad AI/ML hackathon (YUVAAN 2026), out of 7,600+ registrants, with VIVIRITY Intelli-Credit. It reads 500+ page annual reports and turns them into loan risk assessments in under 5 minutes instead of days, on a non-embedding RAG pipeline that costs 98% less in API overhead than vector-based retrieval. Earlier that year I built a drone LiDAR to drainage network pipeline for the MoPR Geospatial Hackathon at IIT Tirupati.

My first IEEE paper, as primary author — Vision-Language Based Real-Time Assistive System for Outdoor Navigation of the Visually Impaired in Indian Urban Environments — was published at ISAECT 2025 in Mohali. I fine-tuned BLIP with a 3-stage LoRA strategy on a custom 427-scene dataset built from Google Street View imagery, for +15.6% across BLEU, METEOR, ROUGE and semantic similarity, and deployed it on a Raspberry Pi through an edge-cloud hybrid architecture. The model is on HuggingFace and the code is open-sourced.

I write about AI on Medium and post explainers on my YouTube channel. Mostly notes from things I actually built and broke: fine-tuning vision-language models on almost no compute, retrieval that survives contact with production, loop engineering. For all the latest I am usually on 𝕏/Twitter or GitHub.

B.Tech in Artificial Intelligence and Data Science at Velagapudi Ramakrishna Siddhartha Engineering College (VRSEC), Vijayawada, at a CGPA of 8.69/10. Deep learning, computer vision and NLP, plus whatever I could teach myself in the gaps — which is where the papers, the hackathons and most of the projects below came from.