Ansh Singhal

• Architected and shipped a production-grade AI security gateway as sole owner across backend, frontend, DevOps, cloud deployment, and scaling — centralized pre/post-LLM security (SHA-256 auth, PII detection, prompt-injection mitigation) across 500+ endpoints, scaled with async concurrency to 100k+ RPS at a 99% reduction in unsafe LLM behavior.
• Drove critical feature expansion on an application-security (DevSecOps) platform — introduced CBOM scanning coverage, shipped AI-powered vulnerability remediations that cut manual triage time by 60%+, and optimized platform performance by 40%+.
• Engineered core features on an endpoint AI-governance solution guarding against sensitive-data exfiltration across 5+ widely-used enterprise AI tools — a lightweight agent + local-proxy architecture serving 1000+ enterprise endpoints.

• Contributed 4 production security analyzers to IntelOwl, an open-source threat-intelligence platform used by security teams worldwide, expanding automated threat-detection coverage by 23% across IP addresses, file hashes, and URLs.
• Added pytest-based validation suites to improve analyzer reliability and regression protection.
• Fixed Django migration and data-consistency issues, improving migration success to 99.9%.
• Resolved Docker networking and caching issues to improve deployment stability and analyzer execution.

• Architected and shipped a production AI security gateway as sole owner across backend, frontend, DevOps, cloud, and scaling — pre/post-LLM security (SHA-256 auth, PII detection, prompt-injection mitigation) fronting 500+ endpoints.
• Scaled with async concurrency from 100 to 100k+ RPS at a 99% reduction in unsafe LLM behavior.

FastAPI, ASGI/Uvicorn, Redis, PostgreSQL, LiteLLM, Presidio, Docker, AWS

• Leading and building from scratch an independent zero-day research engine across 4 ecosystems (npm, MCP servers, Docker images, Hugging Face models), with 50+ custom static-analysis engines (zero open-source libraries) mining 49,522 signals from live registry traffic against 228,020 known-malicious packages.
• Designed a multi-stage LLM triage pipeline at 100% recall and 93% precision, zero false positives on the OSSF malicious-packages corpus, at a 15x cost reduction over baseline.

FastAPI, Next.js, PostgreSQL, Docker, Python

• Built an agentic RAG engine over 10,000+ legal documents at 97% retrieval accuracy, with multi-agent query decomposition, reranking, and deterministic grounding.
• Served citation-traced answers at sub-500ms P95 on AWS containerized microservices.

FastAPI, LangGraph, LangChain, Vector Search, AWS

• Built a real-time voice-streaming backend holding sub-200ms latency for 100+ concurrent sessions, with Kafka consumer groups and Kubernetes HPA backed by PostgreSQL, MongoDB, and Redis.
• Tracked consumer lag and P95 latency to monitor performance across the distributed streaming pipeline.

Kafka, WebSockets, Kubernetes, FastAPI, PostgreSQL, MongoDB, Redis

• Designed a dual-head U-Net translating all-weather SAR radar to optical flood imagery at 1024x1024, using NDWI water-index supervision so flood boundaries are learned as water, not texture.
• Reached FID 28.4 / PSNR 31.25 / SSIM 0.94 with EMA-stabilized adversarial training over ~200 GPU-hours, every run tracked in Weights & Biases.

PyTorch, PyTorch Lightning, CUDA, Docker, Weights & Biases