Dhruv Shah's Work | ContraWork by Dhruv ShahDhruv Shah
AI Specialist Automating Workflows & Building AI Solutions
Hybrid Semantic Search Engine
Keyword search misses results that mean the same thing in different words. Pure vector search misses exact matches.
I built a hybrid search engine that combines both. Text is embedded with the all-MiniLM-L6-v2 sentence model and indexed in FAISS, and each query blends vector similarity with fuzzy text matching (weighted 70/30), so results match on meaning and on exact terms.
Built as a FastAPI backend with a React frontend, packaged with Docker so it runs with one command.
What this shows: I understand how retrieval works under the hood, the same retrieval that powers RAG chatbots, and where it fails. AI Coding Agent & Voice Model Evaluation
Ongoing paid work evaluating frontier AI models for an AI training platform.
For coding agents, I work through the same real-world bug in a codebase with two different agents, then score each on six quality dimensions plus a head-to-head comparison: correctness, reasoning, tool use, and where each one went wrong.
For voice models, I rate AI narration quality against detailed rubrics and review other evaluators' work for accuracy and consistency.
What this shows: I know how leading AI agents actually fail on real tasks, and I can evaluate them rigorously and consistently, which is exactly what teams need before shipping an AI feature. OpenTruth: Independent Verification Protocol
Claims made by AI systems and organisations are hard to check, and the evidence behind them can be quietly changed after the fact.
I built OpenTruth, an independent verification protocol that links each claim to its supporting evidence in a hash-protected evidence graph. Because every piece of evidence is hashed, any later tampering becomes detectable, and anyone can trace a claim back to exactly what supports it.
What this shows: I can design trust and verification systems end to end, from the data model to the integrity guarantees. Sentinel: MCP Security Guard Benchmark
Problem: MCP guard tools claim to block attacks on AI agents, but there was no consistent way to hold them accountable.
What I built: A benchmark with a sandboxed testing harness that runs guard tools against 6 classes of attacks in a controlled environment, so their protection can be measured and compared fairly.
Skills: AI security, MCP, adversarial testing, Python, sandboxing, benchmarking