Risk Management

Hi Mitch,

Thanks for throwing these great questions into the mix. It's the perfect time to look at what's actually happening behind the "black box" of generative AI. LLMs don't understand engineering intent; they simply predict the most statistically probable next token. This explains the exact frustration Bill pointed out-why standard AI often just echoes back what you want to hear. The output is entirely trapped by the immediate context it's fed.

To solve this, I use a Retrieval-Augmented Generation (RAG) architecture within a closed environment, giving the AI access only to a verified reference library. I deploy this natively using Power Apps and Power Automate (AI Builder).

I recently stress-tested this exact low-code framework under extreme constraints in a high-stakes crisis environment. Operating within a strict 72-hour execution window "For more information; kindly see my LinkedIn post , we used it to handle complex, multi-type data influxes, real-time mapping, and automated sorting for thousands of data points. The system held because we tightly controlled the hyperparameters to maintain absolute data integrity:

Temperature (Creativity): Dropped to the lowest quadrant. This forces the AI to be strict and deterministic, sticking 100% to verified reference files instead of being "creative."

Precision (Top_p): Capped below average to restrict wild logical leaps and completely prevent "hallucinations."

Whether you are managing emergency logistics or structural risk assessments, the professional workflow remains identical:

Isolate the Repository: Lock down your clean codes, specific contracts, or standard procedures (your RAG library).

Target the Prompts: Instruct the AI to map and extract risk types, levels, and consequences only from those source files.

Calibrate Thresholds: Lower both temperature and precision settings.

Validate: Run rigorous testing loops inside that closed library before relying on the output.

Recommendation:

Use this setup strictly as a first pass during the initial risk identification phase to stretch your thinking and catch blind spots. It is a powerful assistant for handling massive data, but it is not an autopilot. You absolutely still need a highly experienced team of risk professionals at the wheel to audit the system and make the final, critical action.

Best regards,

------------------------------
Abubakr Gameil, R. ENG, M. ASCE®️,
MSc-Holder, [ SEI, EWRI, CI, ISSMGE ]Mermber
Past / Chairman & Director General
Almanassa Engineering International Co. Ltd,
Khartoum, Sudan
Currently / UAE- Humanitariam Residency
NXN- Central branch -Al Fujairah,
PO.Box : 1142 (Fujairah)
------------------------------

Wow! That sounds impressive, Jameel. The LinkedIn post is intriguing. I hope we can hear more about your team's work here in the CoP.

Bill Mc

------------------------------
William McAnally Ph.D., P.E., BC.CE, BC.NE, F.ASCE
ENGINEER
Columbus MS
------------------------------


Original Message:
Sent: 06-23-2026 04:31 AM
From: Abubakr Gameil
Subject: Application of AI for Risk Assessments

Hi Mitch,

Thanks for throwing these great questions into the mix. It's the perfect time to look at what's actually happening behind the "black box" of generative AI. LLMs don't understand engineering intent; they simply predict the most statistically probable next token. This explains the exact frustration Bill pointed out-why standard AI often just echoes back what you want to hear. The output is entirely trapped by the immediate context it's fed.

To solve this, I use a Retrieval-Augmented Generation (RAG) architecture within a closed environment, giving the AI access only to a verified reference library. I deploy this natively using Power Apps and Power Automate (AI Builder).

I recently stress-tested this exact low-code framework under extreme constraints in a high-stakes crisis environment. Operating within a strict 72-hour execution window "For more information; kindly see my LinkedIn post , we used it to handle complex, multi-type data influxes, real-time mapping, and automated sorting for thousands of data points. The system held because we tightly controlled the hyperparameters to maintain absolute data integrity:

Temperature (Creativity): Dropped to the lowest quadrant. This forces the AI to be strict and deterministic, sticking 100% to verified reference files instead of being "creative."

Precision (Top_p): Capped below average to restrict wild logical leaps and completely prevent "hallucinations."

Whether you are managing emergency logistics or structural risk assessments, the professional workflow remains identical:

Isolate the Repository: Lock down your clean codes, specific contracts, or standard procedures (your RAG library).

Target the Prompts: Instruct the AI to map and extract risk types, levels, and consequences only from those source files.

Calibrate Thresholds: Lower both temperature and precision settings.

Validate: Run rigorous testing loops inside that closed library before relying on the output.

Recommendation:

Use this setup strictly as a first pass during the initial risk identification phase to stretch your thinking and catch blind spots. It is a powerful assistant for handling massive data, but it is not an autopilot. You absolutely still need a highly experienced team of risk professionals at the wheel to audit the system and make the final, critical action.

Best regards,

------------------------------
Abubakr Gameil, R. ENG, M. ASCE®️,
MSc-Holder, [ SEI, EWRI, CI, ISSMGE ]Mermber
Past / Chairman & Director General
Almanassa Engineering International Co. Ltd,
Khartoum, Sudan
Currently / UAE- Humanitariam Residency
NXN- Central branch -Al Fujairah,
PO.Box : 1142 (Fujairah)
------------------------------