hi, i'm Sourasis
a second-year computer science undergrad at JIS College of Engineering, working mostly on machine learning and data science. i build small, honest tools — leak-free prediction pipelines, evidence-first LLM evaluations, and utilities that make model behaviour and API costs legible.
i like the boring parts done well: validation that doesn't leak, features you can explain, and results someone else can reproduce. right now i'm reading broadly across statistical modeling and applied NLP, and shipping whatever i learn next.
building BioAgent-ST · open to internships for fall 2026 · --:--:-- IST
Predicting wine quality from 11 physicochemical measurements, end to end.
Exploratory analysis of the UCI vinho verde dataset — distributions, correlations, duplicates, outliers — then a tuned Random Forest Regressor (MAE 0.414, R² 0.553) against a Linear Regression baseline, with permutation importance surfacing alcohol, sulphates and volatile acidity as the strongest signals.
An evidence-first way to see where RAG systems hallucinate.
A 60-question benchmark across 51 space-mission documents comparing baseline, loose Top-3 and strict Top-3 retrieval. Every answer is split into claims and checked against the retrieved evidence, so failures are visible instead of hidden behind one score: hallucination falls 93.3% → 6.7% and token F1 rises 0.22 → 0.85.
A tokenizer, converter and API cost lab in a single page.
Live token splitting across four exact OpenAI encodings, a token-ID decoder, a both-way words ⇄ tokens converter, and a cost calculator over a price table of 61 real models. The byte-level BPE engine is verified bit-identical to tiktoken on 20 test cases — no framework, no build step, no backend.
Maritime oil-spill forensics from Sentinel-1 SAR — Smart India Hackathon 2026.
Part of Team CODE_FORUN. A custom 1.17M-parameter CNN with Grad-CAM localizes slicks on SAR imagery, while U-Net segmentation, Bonn Agreement volume modeling and a Lagrangian drift hindcast reconstruct a probable discharge origin. PostGIS then correlates AIS trajectories and blackouts to rank investigative leads.
Structured, deterministic feedback on technical CVs.
A shipped app that scores résumés against strict JSON output schemas, keeping parsing deterministic across wildly different formats. Deployed on Vercel with streaming APIs.
Team CODE_FORUN · Smart India Hackathon 2026
SAR forensics pipeline: CNN + Grad-CAM localization, U-Net slick segmentation, Lagrangian drift hindcasting and PostGIS AIS attribution.
nlp / evaluation
Built a 60-question RAG benchmark with claim-level evidence checks and regenerated category, ablation and claim-audit charts from the results run.
live on Vercel
Designed, built and shipped an LLM résumé-feedback app with strict JSON schemas, deterministic scoring and streaming APIs.
JIS College of Engineering · 2nd year
Coursework and side-quests in statistical modeling, validation design and shipping small tools to production.
Statistical modeling · Exploratory data analysis · Feature engineering · XGBoost · LightGBM · CatBoost · scikit-learn · Grouped k-fold CV · Pandas · NumPy
FastAPI · REST APIs · Microservices · Prompt engineering · LLM APIs & streaming
Python · C · SQL · HTML5 / CSS3 · Git / GitHub · Linux / Bash · Streamlit · Vercel
Retrieval misses, bad chunk boundaries, entity confusion, weak verification. TruthScope makes each failure mode visible instead of averaging it away.
explore TruthScope ↗How forcing deterministic scoring schemas tamed heterogeneous résumés — and made streaming deploys painless.
read the code ↗i'm open to internships, research collaborations, and conversations about data science, machine learning, or hard engineering problems. email probably works best — i write back.
