Dhruv
Ladani
Data → models → AI prototypes people say yes to. Built with a PM's eye.
Projects built
DSA problems solved
Research papers
Data & ML
PySpark ETL, feature engineering and ML/DL models on client data, plus Python pipelines that run unattended: validated, deduplicated and versioned in git.
Prototype fast
Client and internal demos built with product teams, from problem definition and KPIs to a working prototype, with a PRD and decision log when the product needs one.
Agentic AI, measured
LangGraph agents, RAG and MCP tooling with evaluation built in: held-out test sets, Ragas and per-stage metrics, so every system ships with numbers, not just a demo.
Tollgate
Teams letting AI agents call real tools get a checkpoint in front of every call — an MCP proxy that allows, holds for a person, or blocks each tool call and scans what tools return for planted instructions. On 229 held-out calls it caught 96.3% of risky calls in a median 321 ms, about 12x faster and 60x cheaper than asking Gemini on every call.
96.3% of risky tool calls caught in 321 ms — 12x faster and 60x cheaper than Gemini judging every call (229-call held-out eval).
LLM Atlas
Developers choosing a model get a leaderboard that says not just who scored highest, but whether the benchmark still means anything and who reported the number — every score joined to its benchmark's health and labelled vendor-reported or independent, refreshed daily from five sources.
Flags saturated and contaminated benchmarks and labels every score vendor-reported or independent — refreshed daily from 5 sources.
Smart Job Tracker
An AI-powered job search pipeline that automates lead generation and evaluation. Solves the fragmentation of job boards by aggregating sources, scoring leads via LLMs, and refining accuracy through feedback loops.
1,247 jobs processed at 87% human–AI scoring correlation → 8 interviews and 2 offers.
Agentic Compliance Auditor
Compliance analysts auditing dense policy documents get cited answers they can check instead of a black-box chatbot — a self-correcting CRAG + Self-RAG pipeline backed by a four-layer evaluation harness, where 98.9% of citations are supported by their source on a 25-question golden set that includes adversarial and unanswerable questions.
99% of citations backed by their source on a 25-question golden set with adversarial cases — and the evals caught its own reranker hurting recall.
Let's build somethingtogether
Currently an Associate Application Engineer (AI/ML) at Impresa.ai. Open to full-time roles in AI/ML engineering, forward-deployed engineering and AI product, and to a small number of freelance builds.