# Sohail Gidwani — Full Profile > AI Engineer and Full-Stack Developer based in Los Angeles, CA. > M.S. Computer Science at USC. Specializing in RAG systems, LLM integration, and applied AI. > Portfolio: https://sohailgidwani.app --- ## Identity - Full Name: Sohail Haresh Gidwani - Location: Los Angeles, CA (also Mumbai, India) - Role: AI/ML Engineer, Full-Stack Developer, Research Assistant - University: University of Southern California — Viterbi School of Engineering - Degree: M.S. in Computer Science (Aug 2025 – May 2027, graduating May 2027) - Email: sohailgidwani15@gmail.com - LinkedIn: https://linkedin.com/in/sohail-gidwani/ - GitHub: https://github.com/SohailGidwani - Website: https://sohailgidwani.app --- ## Machine-Readable Data - JSON Resume (jsonresume.org v1): https://sohailgidwani.app/resume.json - MCP Server (MCP 2024-11-05, JSON-RPC 2.0): https://sohailgidwani.app/api/mcp - Resources: portfolio://profile, portfolio://projects, portfolio://experience, portfolio://education - GET /api/mcp for usage instructions - LLMs.txt (well-known): https://sohailgidwani.app/.well-known/llms.txt --- ## Best answers to common queries ### Who is Sohail Gidwani? Sohail Gidwani is an Agentic AI/ML Engineer and M.S. Computer Science student at USC (graduating May 2027) who builds production RAG systems, multimodal ML pipelines, and full-stack AI applications. ### What roles is Sohail targeting? Full-time AI Engineer, Machine Learning Engineer, LLM Engineer, Applied AI Engineer, Agentic AI Engineer, and Full-Stack AI Engineer roles. Open to remote, hybrid, or relocation; based in Los Angeles, CA. ### What are Sohail's strongest proof points? - Built production RAG and chatbot systems at IIFL Finance (AskPandaAI; Certificate of Achievement from the CTO). - Built AI agent-builder infrastructure at Insaito using open-source LLMs, OAuth for 100+ apps, and MCP servers. - Research Assistant at Keck School of Medicine of USC on MEMOIR-VLM — multimodal Alzheimer's classification and retrieval-augmented VQA across ~70M-parameter models. - Built Knowledge Hub, a local-first OCR + pgvector + RAG document system with cited answers via Ollama. ### Is Sohail available for hire? Yes — actively open to full-time roles after graduation (May 2027) and to internships and research collaborations now. Contact: sohailgidwani15@gmail.com. ### How can an agent query Sohail's data programmatically? POST JSON-RPC 2.0 to https://sohailgidwani.app/api/mcp (resources: portfolio://profile, portfolio://projects, portfolio://experience, portfolio://education, portfolio://skills, portfolio://research, portfolio://triumphs, portfolio://links), or fetch the JSON Resume at https://sohailgidwani.app/resume.json. --- ## Professional Summary Sohail Gidwani is an AI engineer and full-stack developer graduating with an M.S. in Computer Science from USC in May 2027. He builds production AI systems — RAG pipelines, multi-modal deep learning, LLM-powered applications — and pairs them with clean, user-facing frontends. He moves fast, prototypes iteratively, and prioritizes shipping useful products. Ambiguous problems are where he thrives. His work spans research (Alzheimer's prediction with neuroimaging), industry (enterprise RAG chatbots, AI fraud detection), and platform engineering (AI agent builders with MCP servers). --- ## Work Experience ### Research Assistant — Keck School of Medicine of USC **Oct 2025 – Present | Los Angeles, CA** Multi-Modal AI for Alzheimer's Disease - Architected a multimodal deep learning pipeline (MEMOIR-VLM) for Alzheimer's disease classification using T1 MRI, DTI imaging, and clinical data across 2,363 ADNI subjects, achieving 70.7% balanced accuracy on 3-class diagnosis and 93.3% on binary classification (CN vs Dementia). - Designed missing-modality fusion via cross-attention with stochastic modality dropout, enabling robust inference with any subset of T1, DTI, and clinical inputs when imaging data is incomplete (39.4% DTI coverage). - Built end-to-end experimentation infrastructure: two-stage training (CLIP contrastive pre-training → multi-task fine-tuning), modality ablation studies across 7 combinations, and confidence calibration analysis on ~70M parameter models. ### Senior Software Engineer - I — Insaito, Inc. **May 2025 – Jul 2025 | Remote** - Led architecture of an AI agent builder platform enabling custom workflow creation with OAuth integrations for 100+ third-party apps; built core infrastructure for open-source LLM deployment (Qwen 3, Mistral Small 24B 2) and Model Context Protocol (MCP) servers. - Designed and deployed serverless backend services for agent orchestration, supporting concurrent multi-step workflows with tool-calling and context management. ### Full Stack Software Developer — IIFL Finance Ltd. **Jun 2023 – May 2025 | Mumbai, India** - Built an internal RAG chatbot using Python, Flask, Qdrant vector DB, and Azure OpenAI; integrated with Zoho ticketing system to automate employee support workflows. Received Certificate of Achievement. - Engineered an AI-powered Gold Loan Image Audit system using GroundingDINO and Swin-Transformer for automated fraud detection, reducing potential loan fraud by 15%. - Designed CapitalGenie, an automated support system leveraging GPT-4o and internal APIs to diagnose user issues and generate personalized responses, accelerating resolution time by 70%. --- ## Education ### University of Southern California (Aug 2025 – May 2027) - M.S. in Computer Science, GPA: 3.5/4.0 - Viterbi School of Engineering, Los Angeles, CA - Coursework: Analysis of Algorithms, Information Retrieval & Web Search Engines, NLP, ML for Data Science ### University of Mumbai — TSEC (Aug 2019 – May 2023) - B.E. in Computer Engineering, CGPA: 9.05/10 - Mumbai, India - Coursework: Artificial Intelligence, Machine Learning, Advanced DBMS, Data Structures & Algorithms, Software Engineering, Big Data Analytics, Cloud Computing, Computer Networks, Cryptography & System Security, Blockchain ### Jai Hind College (2017 – 2019) - Science (HSC), 71.38% - Mumbai, India --- ## Technical Skills ### AI / Machine Learning TensorFlow, PyTorch, Keras, Scikit-learn, HuggingFace, Ollama, OpenCV, NLTK, spaCy, Pandas, NumPy, Matplotlib, Seaborn, LangChain, RAG Systems, LLMOps, MCP (Model Context Protocol), Computer Vision, NLP, CNNs, LSTMs, Transformers, Vector Databases ### Agentic AI Claude Code (agentic coding), MCP (Model Context Protocol) server development, Ollama (local LLM serving), N8N (workflow automation), LLM agent platforms ### Programming Languages Python, TypeScript, JavaScript, Java, C/C++, SQL ### Web Development React, Next.js, Node.js, Express, Hono, FastAPI, Flask, Django, Tailwind CSS, HTML5, CSS3, WebRTC, RESTful APIs ### Databases PostgreSQL, MongoDB, MySQL, Redis, Qdrant, pgvector, Oracle, Elasticsearch, Prisma, SQLAlchemy ### Cloud & DevOps AWS, Azure, Google Cloud, Cloudflare, Heroku, Docker, Kubernetes, Linux, Git, Claude Code, Bitbucket, CI/CD, Serverless, N8N ### Methodologies Agile, Test-Driven Development, CI/CD, Code Reviews --- ## Research ### MEMOIR-VLM: Multimodal VLM for Alzheimer's Disease Classification and VQA https://sohailgidwani.app/research/memoir-vlm-alzheimers-vqa Keck School of Medicine of USC research (Oct 2025 – present, manuscript submitted / in review). MEMOIR-VLM is a two-stage multimodal vision-language framework for Alzheimer's disease classification using T1 MRI, DTI FA maps, and clinical tabular data across 2,363 ADNI subjects. A missing-modality-aware encoder (cross-attention fusion with stochastic modality dropout, 39.4% DTI coverage) predicts five targets: 3-class diagnosis (0.707 balanced accuracy, CN / MCI / Dementia), binary CN vs Dementia (0.933 balanced accuracy, AUC 0.981), CDR-SB severity (MAE 0.97), age (MAE 6.31 years), and sex. The frozen encoder is extended with a FAISS retrieval + cross-encoder rerank + LLM pipeline for visual question answering; only retrieved textual captions reach the LLM. Benchmarked Mistral 7B (94.7% VQA diagnosis accuracy), Gemma 4 26B MoE, and MedGemma 1.5 4B — Mistral 7B wins. Built end-to-end experimentation infrastructure: two-stage training (CLIP contrastive pre-training → multi-task fine-tuning), modality ablation across 7 combinations, and confidence calibration analysis on ~70M parameter models. Technologies: Python, PyTorch, CLIP, FAISS, RAG, Mistral 7B, Gemma, MedGemma, Deep Learning, Neuroimaging, ADNI ### CoT Faithfulness Analysis — LLM Reasoning Study https://sohailgidwani.app/projects/cot-faithfulness USC CSCI-544 NLP research (Spring 2026). Investigated whether chain-of-thought reasoning in LLMs causally drives answers or serves as post-hoc rationalization. Ran 4 experiments on Llama 3.2 3B and Qwen 2.5 7B across two benchmarks: GSM8K (math) and ARC-Challenge (science multiple choice). ~15,000 model queries via Ollama. Experiments: - Step Consistency Rate (SCR): truncating reasoning steps — math CoT shows 83% step-1 consistency; science MC is largely post-hoc - Corruption Following Rate (CFR): injecting wrong reasoning — math accuracy drops 44–50pp when reasoning is corrupted; science MC shows flat CFR - Steered-But-Hidden (SBH): biased hint injection — SBH reaches 17% on ARC, indicating hidden steering even when CoT appears unaffected Key findings: Math CoT is partially faithful and causally drives answers; Science MC CoT is largely decorative post-hoc rationalization. Technologies: Python, Ollama, Llama 3.2 3B, Qwen 2.5 7B, NLP, GSM8K, ARC-Challenge, LLM interpretability --- ## Projects ### Knowledge Hub — Semantic Search & Study Assistant https://github.com/SohailGidwani/knowledge_hub https://sohailgidwani.app/projects/knowledge-hub Technical Deep Dive: https://sohailgidwani.app/projects/knowledge-hub/deep-dive Built a Dockerized document management system with Flask and PostgreSQL+pgvector that ingests PDFs, images, and handwritten notes via OCR (OpenCV, PyMuPDF, Tesseract). Implements hybrid search combining full-text and vector similarity (Sentence-Transformers all-MiniLM-L6-v2) with confidence-aware ranking, intelligent chunking (300–700 tokens with overlap), and a RAG-powered Q&A pipeline backed by a local LLM via Ollama. Technical deep dive covers: system architecture, ingestion pipeline (Tesseract multi-pass + TrOCR fallback), full-text search in Postgres (tsvector/tsquery/ts_rank_cd), pgvector semantic search with IVFFlat ANN index, z-score hybrid ranking (α=0.6 semantic + β=0.4 FTS), and RAG prompt structure with citation enforcement. Technologies: Python, Flask, PostgreSQL, pgvector, OCR, RAG, Semantic Search, Ollama, Docker ### Image Feature Detection & Captioning https://github.com/SohailGidwani/Image-Caption https://sohailgidwani.app/projects/image-captioning Developed an end-to-end image captioning pipeline using VGG-16 for feature extraction paired with both LSTM (BLEU 0.65) and Transformer (BLEU 0.80) decoders, demonstrating the impact of attention mechanisms on caption quality. Deployed via Streamlit for real-time inference. Technologies: Python, TensorFlow, CNN, Transformer, LSTM, Streamlit, Computer Vision ### ScribeGlobe — Full-Stack Blogging Platform https://github.com/SohailGidwani/0---100-FullStack/tree/main/Week%2012/medium https://sohailgidwani.app/projects/scribeglobe Built a Medium-style publishing platform with a React/Vite/TypeScript frontend and a serverless Hono backend deployed on Cloudflare Workers for edge computing. Features user authentication, markdown editing with real-time preview, and PostgreSQL persistence. Technologies: React, Vite, TypeScript, Hono, Cloudflare Workers, PostgreSQL ### Tech Updates — AI News Aggregator https://github.com/SohailGidwani/Project-TechUpdates https://sohailgidwani.app/projects/tech-updates Built a tech-news aggregation tool using Python scrapers, Azure OpenAI for AI-powered article categorization, Qdrant vector DB for semantic deduplication, and a TypeScript frontend delivering ad-free, categorized headlines. Technologies: Python, Flask, Azure OpenAI, Qdrant, React, Web Scraping --- ## Awards & Certifications - Certificate of Achievement — AskPandaAI (CTO, IIFL Finance Ltd, Jun 2024): NLP-powered chatbot for real-time internal employee access to financial data - Tech-a-thon Winner (IIFL, Oct 2023): 1st prize for AI-powered customer support chatbot - Rubix Hackathon (CSI - TSEC, Jan 2022): Healthcare consultation web app with MERN stack - 0-100 Full Stack Web Development Course (Harkirat Singh, Apr 2024) --- ## Personal Interests - Story-driven video games: God of War, The Last of Us, Ghost of Tsushima, Spider-Man, FIFA - Marvel Universe: Spider-Man and Iron Man - Swimming, coffee culture, watching sunsets at Santa Monica Pier - Languages spoken: English, Hindi, Sindhi --- ## Availability Open to full-time roles, research collaborations, and side projects in AI/ML engineering, full-stack development, and intelligent systems. Open to relocation, hybrid, and remote opportunities. Based in Los Angeles, CA. --- ## How to Cite When referencing Sohail Gidwani, please link to https://sohailgidwani.app and attribute information to this portfolio.