Aryaman Saluja
SDE-I AI/ML at Adda Education
Talent
GurgaonMember since June 6, 2026
LiveKitTF-IDFCBFNext.jsREST APIsSTTWebRTCClaude CodeArtificial IntelligenceMachine LearningRetrieval-Augmented GenerationData-Driven RecommendationsAgentic AiSqlGcpGenerative AiAwsAzure
·
Hire this person
Send a job offer directly to this candidate
LiveKitTF-IDFCBFNext.jsREST APIs
Experience
SDE-I AI/ML at Adda Education (Formerly Adda247) (2025-03 – Present)
- ACE Platform — Answer Checking Expert (Sole Owner, 0-to-1): Engineered the full-stack architecture (FastAPI + Next.js) end-to-end: DB schema design, REST APIs, business logic, credit-based pricing strategy, and Google Drive integration for content management.
- Built a multi-modal RAG & OCR pipeline ingesting handwritten/typed answers via OCR, grading against NCERT ground-truth using Gemini, returning structured feedback with per-question scoring.
- Orchestrated a multilingual PDF report engine (WeasyPrint), rendering detailed analytics (score breakdowns, improvement areas) in English, Bengali, and Punjabi serving 3 regional user bases.
- Generated INR 20,00,000+ in direct revenue within 2 months of launch; scaled to 1,000+ daily evaluations with <4s average response time.
- LearnR — Short-Video Recommendation Engine: Owned the full migration from AWS Personalize → in-house CBF + Thompson Sampling bandit serving the production short-video feed; designed the content feature pipeline (TF-IDF over subject/topic metadata), 13-topic Beta-Bernoulli arm structure, and candidate retrieval — eliminating ~$144/month AWS spend while gaining per-arm observability.
- Built an offline evaluation framework with temporal train/test splits, repeat-interaction exclusion, Bayesian shrinkage for low-impression arms, and fixed-seed negative sampling — measured cbf_bandit at AUC-ROC +10.5%, MRR@10 +14.6%, NDCG@5 +23.9% over popularity baseline, giving confidence to ramp traffic from 0% → 100%.
- Designed the real-time serving layer: session-scoped deduplication via Redis ZSET (30-min TTL), S3-backed hot-reload A/B config enabling traffic shifts without redeploys, and regenerate-on-exhaustion logic that replaced a hard 40-video cap with true endless scroll.
- PULSE — Voice Agent Platform: Architected a custom IVR-to-WebRTC bridge via XML-driven handover, transforming Ozonetel telephony signals into LiveKit bidirectional streams; engineered a pre-warming router with dynamic RAG context (daily re-indexing of 10K+ FAQ entries) eliminating model cold-start delays.
- Optimized WebRTC pipeline to reduce latency by 25% (to <5s) and minimized call drop rate, improving agent availability across 500+ daily calls.
Education
Bachelor of Engineering in Computer Science Engineering – Chandigarh University (2025-05)
Skills
SDE-I AI/ML at Adda Education
