I build machine learning systems and ship them as real products — five-service fraud-scoring platforms, a from-scratch LoRA implementation deployed on serverless GPUs, an RL agent that outperforms PID control on a chemical plant simulator. I spend as much time on the FastAPI layer and the Docker Compose file as on the model itself, and I'd rather report a clean null result honestly than dress up a weak one. Chemical Engineering undergrad at IIT Indore, currently open to ML / SDE internships.
Machine Learning Intern · Suko Labs · Hyderabad (Remote) · Mar 2026 – May 2026
- Built an Indic captioning pipeline on Gemini 2.5 Flash across 7+ regional languages, with a Qdrant-backed RAG cache for viral-traffic news lookups.
- Cut pipeline cost 42% and latency from 20s to 14s via asyncio parallelization and a dual-layer cache (32% hit rate), instrumented with Prometheus.
Machine Learning Research Intern · Charak Centre, IIT Indore · Indore, India · Dec 2025 – Feb 2026
- Built a spatial downscaling pipeline (0.625°×0.5° → 0.1°×0.1°) to resolve PM2.5 micro-gradients invisible at coarse resolution, fusing ERA5/MERRA-2 reanalysis with IMD station data.
- Trained an LSTM-CNN hybrid for hourly pan-India PM2.5 forecasting on 20+ engineered features — R² = 0.78, RMSE = 18 µg/m³.
| CreditRisk "Aegis" |
Real-time transaction fraud-scoring platform — five containerized FastAPI microservices behind a gateway, feeding a Next.js dashboard. Transactions are scored with a gradient-boosted model trained on the 590k-row IEEE-CIS dataset, explained with SHAP, and alerts above a risk threshold get a plain-English narration from a Groq LLM call. Prometheus and Grafana track fraud rate, latency, and alert volume across all nine containers.FastAPI XGBoost SHAP Next.js PostgreSQL Docker Prometheus/Grafana
|
| LinkedIn_LoRA live demo ↗ |
A LoRA adapter for FLUX.1-dev built from scratch — a hand-rolled LoRALayer wrapping nn.Linear, no PEFT dependency — trained with flow-matching to generate LinkedIn-style portraits from a prompt. Captioning and prompt-cleaning run through a Groq-hosted Llama 4 Scout call; inference is served serverless on Modal L40S GPUs behind FastAPI.PyTorch Modal FastAPI React bitsandbytes
|
| FilingAlpha |
Turns SEC 10-K text into trading signals (Loughran-McDonald tone, "Lazy Prices" YoY change, risk-factor delta, Fog readability) and tests them with a leakage-free, point-in-time backtest harness — rank-IC, event-study tercile spreads, and a walk-forward classifier with an explicit max(train) < min(test) assertion. On a 12-company universe every signal comes back statistically insignificant and the classifier lands at a coin-flip — reported straight, since a correctly-measured null result is the actual deliverable, not a manufactured edge.FastAPI PostgreSQL Alembic scikit-learn Docker Prometheus/Grafana
|
| RL_Absorber live demo ↗ |
A PPO-LSTM agent that replaces classical PID control on a simulated MEA CO₂-capture plant, trading off capture rate against reboiler duty in real time. A neural surrogate (trained on ~10k Latin-Hypercube-sampled physics runs) gives ~1000× the speed of the underlying simulator, with a hydraulic flooding limit enforced by projecting the policy's output before it reaches the plant. Holds ~93% mean capture in normal operation vs ~85% for PID, and recovers from disturbances in 4–6 steps vs 18–25.PyTorch Gymnasium RecurrentPPO FastAPI
|
| SpeakPortrait |
An extension of AntGroup's Ditto for audio-driven talking-head animation — one photo and an audio clip in, a photorealistic talking video out, in real time. Added a Next.js frontend, an ngrok-tunneled backend, and Zonos TTS with emotional control and zero-shot voice cloning so the pipeline also accepts plain text.PyTorch Next.js Zonos TTS ONNX/TensorRT
|
| 🏅 | Codeforces Expert — ShlokP06, max rating 1632 |
