Hi, I'm Rohit Challa
Fractional CAIO
I build systems that have to prove they work.
Knowledge graphs, AI security, computer vision and robot autonomy — and the production infrastructure underneath.
What I Do
Most of my work sits where a model meets something that can hurt you if it's wrong — a vehicle, a production database, a security control, a customer's bill.
That constraint shapes how I build. The through-line across all of it: the system carries its own evidence.
AI & Machine Learning
Knowledge graphs, and learning from what is absent
Anti-GraphRAG inverts standard graph retrieval to surface the edges a knowledge graph is missing rather than the ones it holds. Freeze the graph at time T, train ComplEx embeddings, rank absent edges by significance with a contrastive head — then validate against what actually appeared afterward.
0.94 P@100 biomedical · 0.86 P@100 MITRE ATT&CK, both validated post-cutoff
Security & AI Red Teaming
Can your AI be turned against you — and can you survive AI used against you?
Two different tests, run as one practice. Inward: model and provider terms, prompt and context injection, the agent harness, the tool and MCP layer, RAG and PII boundaries, agent privilege, governance. Outward: the AI-enabled kill chain, measured against whether anyone actually detected it.
Mapped to OWASP LLM / Agentic / MCP Top 10, NIST AI RMF and ISO 42001
Hardware & Robotics
Autonomy stacks and embedded systems that have to work outdoors
Uncrewed ground vehicle autonomy — ROS 2 on Jetson Orin, LiDAR and camera fused for perception, ESP32 drivetrain control with a hardware E-stop, and a deterministic scenario harness spanning waypoint following, obstacle approach, GNSS-denied navigation and link-loss return-to-home.
Simulation refuses to emit quotable results while any vehicle parameter is unmeasured
Computer Vision & Remote Sensing
Pixels to decisions — from orbit, from a robot, and from a production line
Satellite and aerial imagery turned into decision products for environmental monitoring: fuel-layer characterization, hazard and spread modelling, and risk surfaces built for the people who act on them rather than for a leaderboard. On the robotics side, LiDAR and camera fusion for obstacle detection, with a calibration step that has to run before any range figure is allowed to mean anything. Also inspection and quality grading on moving production lines.
Detection benchmarked with evidence-grade logging — calibrate first, or the range number is void
Software & Product
The whole system, not a notebook
Production Next.js on Vercel, Postgres with row-level security designed before the first insert, Stripe billing with verified webhooks, containerized FastAPI services, and the operational work that decides whether any of it survives contact with users.
Multi-tenant platforms, embeddable widgets with usage metering, CV pipelines end to end
GitHub Activity
Most of the work is in private repositories — this is the part that shows
Need something else? I adapt quickly to your stack. Whether it's legacy code, cutting-edge frameworks, or proprietary systems—I get up to speed fast.