Looking for a job

Srijan Ratrey

@srijanratrey

Message

AI/ML Intern at Kavana who improved NSFW catch rate to 77% and shipped context-aware speech and image generation.

India

Message

What I'm looking for

I'm looking for AI/ML roles where I can keep learning fast and shipping, especially work involving LLMs, evaluation, and taking models from research into production.

At Kavana, I improved the safety classifier’s NSFW catch rate from 66% to 77% across three generations, including a FAISS-based classifier and contributions to a neural-network scorer. I also curated data and analyzed misclassifications to support auto-training and deployment.

I designed an LLM-as-judge evaluation pipeline for D0 conversations, reaching 87% NSFW detection precision and evaluating 1,000 conversations in under 15 minutes. The work boosted user engagement by approximately 6%.

At Kavana, I also delivered context-aware speech-to-text with end-to-end response under 1.5 seconds and shipped real-time image generation that preserved narrative continuity across sessions. Earlier, at Seekho, I developed a hybrid autoplay recommendation engine that increased video completion rates by 15%.

Experience

Work history, roles, and key accomplishments

ML Intern

Kavana

Dec 2025 - Jun 2026 (6 months)

Raised NSFW catch rate 66%→77% by rebuilding the safety classifier over 3 versions (FAISS V2, neural-net V3, auto-deploy). Built an LLM-as-judge pipeline: 87% NSFW precision, ~6% engagement lift, 1,000 chats scored in <15 min daily. Shipped <1.5s speech-to-text after benchmarking 6 vendors on WER and latency. Launched in-chat image generation with cross-session narrative continuity.

AI/ML Intern

Seekho

May 2024 - Aug 2024 (3 months)

Built a hybrid autoplay recommender combining collaborative filtering (watch history) with content-based filtering (video embeddings), lifting video completion 6%. Analysed user-event logs, cutting bounce rate 4%. Ran end-to-end A/B tests on autoplay variants, raising session duration 8%. Benchmarked 3 models (incl. matrix factorization) offline and picked the best, improving accuracy 5%.