About

I'm Srijan Ratrey, an engineer in Bengaluru working on machine learning systems — mostly content safety, retrieval, and the evaluation infrastructure around them.

The thread running through my work is measurement. It is easy to build something that produces plausible output and much harder to know whether it is right. A multi-label classifier that predicts no labels at all scores 89.83% exact-match accuracy on the Jigsaw toxic comment data, at a macro-F1 of exactly zero. A RAG system will happily cite the wrong PDF with total confidence. So a lot of the work I find worth doing is picking a metric that can actually be wrong, calibrating the operating point instead of defaulting to 0.5, and writing down precisely where the thing breaks.

That last part matters to me. The classifier I trained fails silently on Hinglish — it scores a Hindi death threat at 0.000 and fires no labels at all. That is worse than degrading, and it belongs in the README rather than in a footnote, so it is in the README.

Experience

  1. AI/ML Intern

    Kavana · Bengaluru, India

    Dec 2025 — Jun 2026
    • Raised NSFW catch rate from 66% to 77% by re-architecting the safety classifier across three generations — rebuilt V2 from scratch as a FAISS-based 3-label classifier (clean/borderline/explicit), then drove V3, a neural-net scorer, via misclassification analysis, data curation, and auto-training and auto-deployment to production.
    • Built an LLM-as-judge eval pipeline scoring conversations across 5 risk tags (NSFW, abuse, flirting, emotional support, PII) — reached 87% NSFW precision and ~6% engagement lift; benchmarked multiple LLMs and tuned batching to score 1,000 conversations in under 15 minutes via a daily cron.
    • Shipped context-aware speech-to-text at under 1.5s end-to-end by benchmarking 6 vendors (Sarvam, Gemini, OpenAI, ElevenLabs, Cartesia, Smallest AI) on a 100-sample set against a WER and latency trade-off.
    • Shipped real-time, context-aware image generation in-chat — bots generate and persist visuals that maintain narrative continuity across sessions.

  2. AI/ML Intern

    Seekho · Bengaluru, India

    May 2024 — Jul 2024
    • Developed a hybrid autoplay recommendation engine combining collaborative filtering on user watch history with content-based filtering on video embeddings, contributing to a 6% lift in video completion rates.
    • Performed extensive data wrangling and analysis on user-event logs, directly contributing to a 4% reduction in bounce rate.
    • Executed end-to-end A/B experimentation on autoplay variants, resulting in an 8% increase in session duration.
    • Benchmarked multiple approaches — matrix factorization, similarity scoring, embedding optimization — and selected the top performer on offline evaluation metrics.

Education

B.Tech, Data Science and Artificial Intelligence

IIIT Naya Raipur

2021 — 2025

Skills

Languages
Python, SQL
ML & DL
PyTorch, TensorFlow, Keras, XGBoost, CNNs, GANs, YOLO, ResNet, OpenCV
LLMs & NLP
Transformers, LangChain, fine-tuning, LoRA, prompt engineering, GPT-4, Gemini, Claude, T5
Infra & MLOps
Django, Docker, Flask, FAISS, Langfuse, A/B testing, MongoDB, GCP, BigQuery, Vertex AI, Amazon S3

Certificates

Elsewhere