I'm Saud — an AI/ML engineer who likes hard, load-bearing problems: retrieval that can't hallucinate, risk engines that can't miss, vision that runs where there's no internet. Most of what's below started as a hackathon sprint and got hardened into a system — eval harness, ablations, latency budget and all.
- 🧠 Applied AI — hybrid RAG (dense + sparse + cross-encoder rerankers), deterministic risk engines, agentic tool-calling
- 👁️ Vision at the edge — TFLite / MobileFaceNet fully on-device, YOLOv8 in the loop
- 🔩 The unglamorous parts — eval harnesses, ablation studies, property-based tests, offline reproducibility
- 🛡️ Upstream OSS — memory-safety & validation patches merged into
libheifandassimp, one in review atlibavif - 🏆 NIRMAN 2026 Winner — CrackWatch, an AI infrastructure-damage command center (₹75K prize)
- 🎓 B.Tech Computer Engineering ('27) · Thane, India · open to AI/ML roles & collabs
Data → Model → API → Interface — I own the whole pipe.
|
Compound-risk intelligence for industrial safety. A deterministic engine fuses gas sensors, hot-work permits & CCTV into one risk score — LangGraph agents explain the decision, they never make it. Replayed against real disasters (Texas City '05, Jaipur '09), it fires minutes before any single sensor.
Python · FastAPI · LangGraph · NetworkX · YOLOv8 · Gemini
|
Plain English in → Bureau of Indian Standards codes out. Four-stage hybrid retrieval: FAISS (bge-m3) + BM25 → reciprocal-rank fusion → cross-encoder rerank, with a whitelist guard so it can't hallucinate a standard. Auto-clamps to available VRAM, all the way down to CPU.
Python · FAISS · bge-m3 · BM25 · cross-encoder · FastAPI
|
|
Face recognition & liveness detection that works where the internet doesn't. MobileFaceNet on TensorFlow Lite, entirely on-device — built for remote sites with zero connectivity.
React Native · TensorFlow Lite · MobileFaceNet
|
AI command center for civic infrastructure — detects and triages structural damage from imagery, then routes it like an ops platform. Took home NIRMAN 2026. Since rebuilt AI-native on the Lemma SDK →
Vision AI · JavaScript · Lemma SDK
|
|
Real-time Q&A teleprompter: live voice capture → Groq Whisper transcription → Gemini answers, streamed to a glanceable overlay while you keep talking.
Python · Groq Whisper · Gemini
|
sunpath-solar — an agentic AI solar sales rep that qualifies, quotes and books surveys via tool-calling with generative UI. sbi-saarthi — a multi-agent banking concierge, built for SBI Hackathon @ GFF 2026.
TypeScript · Next.js · LLM orchestration
|
Also on the bench: FraudShield (fintech fraud detection) · trader-performance-vs-market-sentiment (211k-trade sentiment analysis) · ride-a-go (Flutter group-riding app) · WMS · more →
Security-hardening patches in C/C++ libraries that ship inside browsers, games and image pipelines:
| PR | What it hardens | |
|---|---|---|
| strukturag/libheif #1840 | Enforces security limits on icef unit allocation — no unbounded memory from crafted HEIF files |
|
| assimp/assimp #6705 | Validates B3D chunk sizes against buffer & parent chunk — stops oversized allocations from malformed models | |
| AOMediaCodec/libavif #3272 | Adopts the __counted_by bounds-safety model on core data buffers |
AI & systems
Product & apps
Tools