Hi, I'm Siddhesh More 👋
I work at the intersection of AI research and engineering.
MS student at ASU building RAG systems, agentic pipelines, and products people actually use.
About
The question I keep coming back to: how do you make a language model less confidently wrong? I work on that as a Research Aide at ASU (4.0 GPA), and I also ship products that 3,000+ people use.
The research side covers cross-model uncertainty quantification, dense retrieval that stays robust on multi-hop queries, and Text-to-SQL systems that beat GPT-4 using open-weight models. Our SNIPER SQL entry ranked #1 worldwide on the BIRD-CRITIC leaderboard. Six papers so far, four accepted or published.
The builder side: Referrlyy has 2,200+ users finding job referrals. BoltPrep has 1,000+ prepping for interviews. Won two hackathons building for nonprofits in under 36 hours. If it can be shipped, I will ship it.
Currently in Tempe, AZ, where the sun is a personal enemy and I still wear a hoodie indoors. Always reading something, always building something.
Work Experience
Architected a production RAG pipeline with LangChain and FAISS on AWS Lambda, serving 500+ concurrent academic queries daily with zero downtime via GitHub Actions CI/CD.
Engineered a Redis caching layer with PostgreSQL connection pooling that cut query latency 60% (800ms to 320ms), now the system's primary performance lever.
Built FastAPI microservices with stateful chat-history retrieval, enabling persistent conversation context across sessions for an AI academic assistant.
Shipped a Flutter and Supabase mobile app from zero to 300+ rural Indian users, with OAuth 2.0 (Google Sign-In), real-time sync, and offline-first UX.
Built full multi-language support (English, Hindi, Marathi) and voice navigation using Flutter TTS/STT, making the app usable without English literacy.
Redesigned the onboarding flow, cutting completion time from 8 to 5 minutes for users unfamiliar with smartphones.
Built RESTful APIs for an ERP system using Django REST Framework, sustaining 50K+ daily requests at sub-100ms average latency.
Architected a Redis caching and Celery async task queue layer, decoupling CPU-heavy report jobs from the main request cycle.
Diagnosed N+1 query patterns in PostgreSQL and added composite indices, cutting report generation from 45 seconds to 6 seconds.
Skills
Languages
Frameworks
ML / AI
Infrastructure
Selected Work
Research tools, shipped products, and hackathon builds.
Research
FluxServe: LLM Inference Server
Continuous batching cut median TTFT from 0.63s to 0.03s; CUDA graphs made decode steps 2.2-6.3x faster
Research
VectorForge: Vector Database in C++17
Hand-written HNSW at 0.83-1.55x FAISS speed at equal recall, and 0 failures across 1,400 kill -9 crash cycles
Research
FilingLens: Cited Q&A over SEC Filings
Cross-encoder reranking lifted text-question accuracy from 71% to 91-93%, with abstention when the answer is not in the filings
More projects
+23% Recall@10 from a LoRA-tuned CLIP over 44K images, 14x faster ANN via HNSW
Hybrid retrieval + rerank (NDCG@5 0.96) alongside a GPT built from raw PyTorch and SFT to DPO alignment
Multi-agent LangGraph system with human-in-the-loop approval and 90% tool-call accuracy on a CLEAR eval
+167% NDCG@10 from a LightGBM ranking stage over two-tower retrieval, with explicit cold-start handling
Tool-using agent scored 76% on 100 held-out cases and stayed at 70% on 6,000-transaction accounts where stuffing the prompt fails outright
A LoRA-tuned 1.5B router reached 98.5% on held-out complaints, and a load test caught a bug that capped the gateway at 18 req/s
Nine LLM judges audited on exact answers: wrong answers accepted ranged from 0% to 91.7%, and thinking mode lifted Qwen3 30B from 87.1% to 96.7%
18,480 runs: undefended models obey 16.5% of planted instructions, a screening model stops all of them, and a planted false fact still gets through
Publications
Papers on making language models more reliable, four of them accepted or published. Tap a title to read the findings.
Blog
View all
Why a Portfolio Website Still Matters
In a world of noisy profiles and AI-generated resumes, your personal website is the one space you fully control and that makes it more valuable than ever.
May 20, 2025
How Saying "Yes" After a Loss Led to My Most Impactful Launch Yet
From hackathon disappointment to a thriving job referral app — discover how a simple decision sparked something bigger than I imagined.
Apr 19, 2025
From Hackathon Rejection to 500+ Downloads: The Story of AquaTrace
This is the story of how a hackathon rejection turned into 500+ downloads — and why sharing what you build matters just as much as building it.
Apr 18, 2025
The Problem That Started It All: Building BoltPrep
How we built an AI-powered mock interview platform in just one week and got 50+ users in two weeks.
Mar 30, 2025