Home / Insights
Insights
Short explainers, the vocabulary we use, and how to get the full build guide.
Most automation either alerts you or acts blindly. A self-healing agent runs a fixed loop: detect, diagnose, plan, fix, verify, report. The verify step matters most. After a fix, the agent checks the data again; anything still wrong is retried or left open and reported. You get a record of what was found, what was fixed and proven fixed, and what is waiting for you.
Anyone can claim an AI agent is accurate. Before every demo our pipeline adds problems to public and synthetic data, such as blank amounts, impossible ages, invalid PAN numbers and double bookings, and records exactly where they are. After the run, the dashboard compares what the agent found with what was planted. In a paid audit we do the same with records your team labels.
Every proposed fix has a risk level. Low-risk fixes are reversible data repairs, such as removing a duplicate. Medium-risk fixes change a business record or send a message. High-risk fixes move money, block access, merge identities or change prices below cost. During a pilot only low-risk fixes run on their own. High-risk fixes always wait for a named person.
The full demo platform runs on one Graviton EC2 server with S3, Bedrock, CloudFront and CloudWatch for about $23 a month at list prices. Bedrock lets us pick the right model per client without storing API keys, and Indian client data can stay in the Mumbai region.
Language models write diagnoses and reports. If a model is slow, rate-limited or not allowed for your data, the platform switches to summaries built from the agent's own findings. Decisions never depended on the model: detection is rules and trained models, and actions follow the risk policy.
A process map, a data contract for the tables involved, the agent's accuracy on 50–100 records your team labels, an ROI baseline in your numbers, and a fixed quote for a pilot or build. The fee is credited if you go ahead within 60 days.
A 386-page engineering guide: AWS account and cost guardrails, data engineering with Kaggle and public data, seven ML models, the LangGraph loop and CrewAI crews, Amazon Bedrock, all 38 agents with code, the Control Tower, n8n workflows, tests and launch with CloudFormation.
Shared with technical evaluators during an audit or partnership discussion.
| Term | Meaning |
|---|---|
| Self-healing agent | An agent that detects problems, fixes the safe ones, verifies its fixes and asks a person about the rest. |
| Planted problem | A fault deliberately added to demo data and recorded, so accuracy can be measured. |
| Detection recall | The share of planted problems an agent found. 1.00 means all of them. |
| Risk policy | Rules that decide which fixes run automatically and which need approval. |
| Control Tower | The dashboard showing every run, issue, fix and approval. |
| Lakehouse | Data organised in layers: raw, bronze (typed), silver (healed), gold (business tables). |
| Amazon Bedrock | The AWS service that gives access to language models such as Amazon Nova and Anthropic Claude. |
| Agent Pack | A bundle of agents for one part of a business, sold as one subscription. |
Watch an agent find planted problems, fix the safe ones and wait for your approval on the rest. Thirty minutes, live.
