Hi, I'm Praneeth — a Master's student in Computer Science (MSCS) at Northwestern University, with concentration in Artificial Intelligence (expected Dec 2026).
I am broadly interested in Machine Learning, Foundation models, and Generative AI, particularly diffusion models. I’m exploring these areas from both theoretical and a application perspective, focusing on developing models that are efficient, robust, and deployable in real-world environments.
Most recently, I interned at Kilwa Technologies LLC as an Machine Learning Engineer Intern, where I built hybrid time-series forecasting model for emerging african markets by combining macro-economic indicators and news-based sentiment signals. Prior to grad school, I was a Software Engineer at Arcesium (D.E. Shaw Group), in the Infrastructure team building distributed job-execution systems, enterprise alerting systems and low-latency data-ingestion pipelines.
I'm currently looking for full-time roles in Machine Learning and Applied AI starting Jan 2027 — if you're working on something interesting, I'd love to connect. Feel free to reach out via email, LinkedIn, or GitHub, or check out my CV.
Fine-Tuning LLMs for Math Reasoning While Preserving Safety Alignment GitHub
Can you make an LLM better at math without breaking its safety guardrails? I fine-tuned Qwen2.5 on GSM8K using LoRA and found that with the right setup, math accuracy jumps from 38% to 81% while the model still scores 88% on safety benchmarks. The trickier part was understanding when and why catastrophic forgetting kicks in — I ran ablations across 10 configurations to get a clearer picture.
Post-Training Quantization (PTQ) for Diffusion Transformers GitHub
Diffusion models are powerful but expensive to run. I looked at whether floating-point or integer quantization holds up better when you aggressively compress PixArt-α to 4-bit weights and 8-bit activations. Turns out FP formats preserve visual quality noticeably better — FID dropped from 42.4 to 38.8 — while keeping CLIP and ImageReward scores on par.
Deformable Object Manipulation with Vision-Language-Action Policies GitHub
Teaching a robot to fold clothes is surprisingly hard — fabric doesn't behave predictably, and small errors compound quickly. I trained SmolVLA from teleoperated demos and got it to a 69% success rate on garment folding, beating both Diffusion Policy (41%) and ACT (61%). To boost generalization, we also trained it on synthetic data generated using NVIDIA's Cosmos-Transfer.
Rotation Invariant Multi-Object Detector GitHub
Standard object detectors quietly assume images are right-side up — flip or rotate a photo and accuracy tanks. I built a pre-processing step using eigenvector analysis that corrects image orientation before passing it to YOLOv3, with no retraining required. On Pascal VOC images rotated 90°–270°, this alone gave a 43% accuracy boost over a ResNet50 baseline.
Deep RL for Real-Time Bidding in Sponsored Search — Literature Review GitHub
Online ad auctions happen in milliseconds, and traditional rule-based bidders struggle to adapt to shifting market dynamics. This review surveys how Deep RL — specifically DQN variants — reframes real-time bidding as a constrained MDP and learns smarter budget-pacing strategies. The best approaches show up to 120% ROI gains over classical methods, though real-world deployment still has open challenges worth digging into.