AI/ML Services
At Radix, we build supervised/unsupervised models to forecast demand, detect anomalies, and classify customers; using NLP, Image Recognition & Deep Learning we solve complex text and related prediction problems; Generative AI with retrieval and guardrails for chat over documents; and Reinforcement Learning/bandits for continuous improvement provide human-like interfaces to client systems. Everything ships with robust MLOps for experiment tracking, model registry, CI/CD, & monitoring - so models run as reliable APIs & plug seamlessly into existing production systems.
Classification/regression, segmentation, anomaly detection using the right one of the plethora of techniques along with explainability is our forte.
Next-best-action, beginning with offline/batch RL and moving to safe online learning with guardrails and human-in-the-loop where decisions are sensitive.
Detection, tracking, OCR; real-time occupancy and queue analytics; edge deployments for low-latency/private environments.
Experiment tracking, model registry, CI/CD, canary/blue-green rollouts, performance & drift monitoring; rollback playbooks.
Chatbot for documents that answers questions with citations, follows access rules, and keeps data private. Light-weight adaptation (LoRA/PEFT) and preference optimization (DPO/RLHF) to fit the domain and tone; agent workflows that call tools/APIs and return structured outputs, with options to run on-prem or in a private cloud.
What you get
- Model cards with metrics, bias/variance notes, and operating thresholds
- Inference services (APIs/streams) with SLOs and monitoring
- Runbooks for retraining, rollback, and incident response
Representative case snapshots
The client required to find relations between Address and Entity, using ML, to help standardize messy data by recognizing when two entries refer to the same place or organization. The solution used two-stage ML approach. It was designed to plug into existing systems via APIs and scales to large databases for deduplication. Radix developed a solution providing recall of 87% for addresses and 89% for entities.
An edTech test preparation company, having a large MCQ question bank, wanted to provide targeted tests to its students. The challenge was that the text contained lots of mathematical and chemical symbols, structures, equations and diagrams etc. Thus, it became a complex text- and image-based classification problem. Radix developed a solution pipeline with 92% accuracy for about 30 topics in each subject.
A major insurance company in Singapore wanted to deploy an automated Fraud Scoring Tool for its travel insurance portfolio. Along with the usual low incident rates in the data for modelling fraud score, interesting challenges like the nature of claims, status of formal police complaint had to be handled. Radix provided a model which with a very high recall and substantial precision which reduced manual investigation by 60%.
What we build, how it performs - Explore our work!
Would love to hear your thoughts!
