About
Articles by Biniam (Bini)
Activity
7K followers
-
Biniam (Bini) Teklehaimanot, MBA, CRISC shared thisThat cloud run I mentioned last week produced real results worth talking about. Melanoma caught early has a 99% five-year survival rate. Caught late, that number drops to 35%. That gap is not a medical mystery. It is a detection problem. I trained EfficientNet-B3 on the HAM10000 dermatology dataset to classify seven types of skin lesions. The model hit 84% melanoma recall. For every 100 melanoma cases in the test set, it flagged 84 of them. Overall accuracy was 63%. That number looks bad until you understand why it is designed that way. Accuracy rewards the model for getting the easy cases right. HAM10000 is heavily skewed toward benign lesions. A model that ignores melanoma entirely can still score above 60% accuracy. That is not a screening tool. That is a false sense of safety. We optimized for recall instead. The tradeoff is real. Melanoma precision sits at 0.28, meaning the model generates false positives. But in a screening context, a false positive sends a patient to a dermatologist. A false negative sends a melanoma home undetected. The cost of an unnecessary biopsy is measured in dollars and anxiety. The cost of a missed melanoma is measured in years. This is not a production system. It is a research prototype on a public dataset. But the decision framework behind it, choosing the right metric for the clinical stakes, is the same decision a HealthTech team building real diagnostic AI has to make on day one. Next step is Vision Transformer. The architecture question is interesting. The metric question is already answered.
-
Biniam (Bini) Teklehaimanot, MBA, CRISC shared thisFollowing up on my last post about that initial 100% failure rate in my melanoma detection model. The Lesson: Small scale can be a liar. When I was running initial tests on a local, underpowered GPU, I worked with a heavily truncated sample size just to get the code to execute. The results were misleading because the model did not have enough data to learn the nuances of rare edge cases. It was a sandbox version that could not reflect real world clinical risk. The Reality Check: Moving the full training to the cloud was the only way to see the truth. Using RunPod, I was able to scale up the compute and process the entire dataset. That is where the real work began. Sometimes, experimenting at a small scale is not just "not enough." It is actually dangerous. It can give you a false sense of security or, in my case, a baseline that looks like a total failure because the sample size is too small to capture the minority class. The Wallet Conversation: I often get asked about the cost of cloud compute. The reality is that the ROI of cloud training far outweighs the "savings" of local experimentation. When you factor in the cost of engineer time spent waiting on slow local runs and the massive business risk of deploying a model built on incomplete data, the cloud is the cheaper option. Paying for high performance compute like RunPod is an investment in accuracy and speed to market. Saving a few dollars on compute while losing weeks of progress or missing critical edge cases is a bad trade for any P&L. Key Takeaways: - Infrastructure matters: Do not let hardware limitations dictate your model logic. - Sample size integrity: Rare cases like melanoma require the full dataset to even show up in the metrics. - ROI over expense: Cloud scalability allows for the iteration speed needed to move past baseline failures and into actual tuning. Are you making decisions based on local "mini-models," or are you looking at the full scale of your data and its actual impact on your ROI?
-
Biniam (Bini) Teklehaimanot, MBA, CRISC shared thisMy model just missed every single melanoma case on the first pass. That is zero percent recall for the most dangerous form of skin cancer. The scary part isn't the number. It's that the overall accuracy looked fine. It was in the sixty percent range. If I was just looking at that high level metric, I might have thought we were making decent progress. I wasn't, though. The system was failing completely at its actual job: helping doctors catch the dangerous edge cases that matter most. We have now established a full pipeline for this high-stakes decision system, which is the real win this week. I don't care about the initial accuracy number or what library we used. I care that we can now reliably measure exactly how badly the model is missing the true positive cases. Now we can start tuning it based on real clinical risk, not convenient vanity metrics. This is how we actually build something that can support better triage and improve early detection workflows, reducing the chance of a critical missed diagnosis. Ask yourself: is your team celebrating model accuracy, or are they measuring the specific business metric where failure costs the most?
-
Biniam (Bini) Teklehaimanot, MBA, CRISC shared thisI used to think my value was tied to the complexity of my code. I spend time building the most intricate models I could imagine. Then I sat across from a leader who had to decide where to put ten million dollars. They didn't care about the model's accuracy score. They asked me if we were going to lose money or make it. I realized we, as the development team, were speaking a language that didn't pay the bills. In 2026, the real win isn't being the best coder in the room. It's being the person who can translate a confidence interval into a budget variance. If you can't explain how a high-stakes decision system protects the wallet, the system won't get built. Technical vanity doesn't move the needle. Business outcomes do. I'd rather hire a business expert who knows a little Python than a Python expert who knows zero business. One person builds a tool. The other architects a solution that grows the P&L. To lead in this space, you have to stop being the person who just executes the script. You have to become the bridge between the data and the revenue. I challenge you to look at your current project. Can you explain its value to a stakeholder without using a single technical term?
-
Biniam (Bini) Teklehaimanot, MBA, CRISC shared thisA 99 percent accuracy rate in fraud detection is often a failure. A machine learning model can be almost perfectly accurate and still threaten the business. I was once tasked with building a high-stakes risk decisioning system. The model we developed was brilliant at finding almost every fraudulent transaction. Our initial success metric looked fantastic. Then we missed one massive fraudulent transaction that cost the company a significant amount. That's when it clicked. Not all errors carry the same price tag. Annoying a good customer with a false positive is frustrating. But missing a deep fake or a sophisticated fraud attack is catastrophic. Technical vanity metrics like standard accuracy don't reflect that cost asymmetry. Risk modeling isn't just math, it's about choosing the business impact you can survive. My job is deciding which "cost of being wrong" the company is actually willing to afford. I challenge you to look beyond high-level accuracy scores in your key decision systems. Ask yourself, what specific type of error will actually show up on your P&L.
-
Biniam (Bini) Teklehaimanot, MBA, CRISC shared thisIn 2026, 80% of AI projects are failing twice as often as traditional IT: but you won't see it on this quarter's dashboard. I spent a morning last week looking at a set of speed sensors. They weren't for a new software build. They were for a car. In F1, if a sensor fails to feed the right data to the driver, the race is over before the first turn. There is no room for vanity. The only thing that matters is the gap between the car and the finish line. I've noticed a pattern with leadership teams lately. They're racing to install the newest AI engines without checking if the sensors actually connect to the wheels. We call these AI initiatives. I call them the invisible failure. It starts with a quiet accumulation of wrong answers. Then comes the degraded trust from the people on the front lines. Finally, the high API bills show up. By the time these costs hit the P&L, it's six months after the big launch party. The wallet is thinner, but the needle hasn't moved. I've learned that building high-stakes decision systems isn't about the tech. It's about the bridge between the code and the cash. My first thought used to be about model fit. That was wrong. Now, I don't care how sleek the engine looks if I don't know where it's going. Before you write a single line of code, you need to find your metric. If you can't name the dollar value of a 5% increase in accuracy, you aren't ready to build. Speed is only a win if you're headed toward growth. Ask yourself: If we turned this system off tomorrow, which line on the P&L would actually change?
-
Biniam (Bini) Teklehaimanot, MBA, CRISC shared thisOperational Impact A leader once asked me how I knew our model wouldn't crumble the moment it hit the real world. It is a fair question. Most systems look like athletes in the lab and collapse on the field. We ran our dermatological AI against a completely held-out test set. The result was 71 percent accuracy. The most important part of that number was the lack of drift. It matched our validation scores almost perfectly. That consistency is what allows a business to scale. It means the model is robust enough to handle the mess of reality. But we didn't stop at the broad average. We looked at the cost of being wrong. Missing a melanoma diagnosis is a massive institutional liability. It is a failure of the system's primary purpose. We prioritized high recall for the most dangerous conditions. We traded some overall precision to ensure we caught the cases that matter most. Reliability isn't just about high numbers. It is about knowing exactly where your system might fail and making sure that failure is affordable. Ask yourself if your technical team is solving for the highest math score or the lowest business risk.
-
Biniam (Bini) Teklehaimanot, MBA, CRISC shared thisEarlier this week, I shared my thoughts on how we hire for skills that AI is quickly absorbing. Data science isn't going away. The math, the models, and the foundation of the craft are still the engine. What is changing is the wrapper around it. The best data scientists I know have always been curious about how the business works. We just never bothered to put that on the job description. I've seen leaders try to overcorrect lately. They want to throw out technical depth and replace it with soft skills. That's a mistake that will cost the wallet later. Don't hire for one or the other. Hire for the bridge. Find the person who has the technical depth and can also sit across from an executive to explain what a number means for next quarter. High stakes decision systems need both to survive. The companies that get this right won't just have better models on a screen. They will make faster decisions with higher confidence. That is a win that compounds every single month. I challenge you to stop hiring technical people who can only talk to other technical people. Ask yourself if your next hire can explain a risk model in a way that actually moves the needle for the P&L.
-
Biniam (Bini) Teklehaimanot, MBA, CRISC shared thisI looked at a job description for a lead data role yesterday. It felt like reading a museum catalog. The list was long. It wanted coding speed. It wanted math fluency. It wanted someone to build the engine from scratch. My first thought was that this person would be a hero. But I was wrong. Most of those technical tasks are being eaten by the tools we just bought. We are still hiring for the work that AI is about to do better than any of us. I see companies spending months looking for a specific technical flex. Meanwhile, the actual business is waiting for an answer. The skills that are depreciating the fastest are the ones we put at the top of the pile. We forget that the real needle moves when someone can sit in a room and understand why the wallet is closing. The most expensive hire is the one who builds a perfect model for a problem that doesn't affect the P&L. I focus on high stakes decision systems because the math has to serve the outcome. If the system is robust but the human behind it can't connect it to revenue, it is just an expensive hobby. Speed matters. But the speed of the code is nothing compared to the speed of a good decision. We need more people who can act as the bridge. People who care more about the win than the tools they used to get there. I challenge you to look at your latest open role. If a machine can do seventy percent of the requirements by next year, what are you actually hiring for.
-
Biniam (Bini) Teklehaimanot, MBA, CRISC liked thisBiniam (Bini) Teklehaimanot, MBA, CRISC liked thisYou will have a reason to smile again #christianity #jesus #christian #bible #God #faith #jesuschrist #christ #prayer #gospel
Experience & Education
Licenses & Certifications
Volunteer Experience
Projects
-
Detecting Accounting Fraud in Public Company Financials
-
Built a deep learning and ensemble-based machine learning framework to detect accounting fraud in publicly traded companies using structured financial reporting data. Applied advanced preprocessing, feature selection, and data balancing techniques to improve prediction of financial misstatements.
Key outcomes included identifying the most influential financial indicators associated with fraudulent reporting and developing a generalizable model capable of flagging risk across diverse…Built a deep learning and ensemble-based machine learning framework to detect accounting fraud in publicly traded companies using structured financial reporting data. Applied advanced preprocessing, feature selection, and data balancing techniques to improve prediction of financial misstatements.
Key outcomes included identifying the most influential financial indicators associated with fraudulent reporting and developing a generalizable model capable of flagging risk across diverse accounting ecosystems.
Tools: Python, TensorFlow, CatBoost, scikit-learn, Git, Jupyter Lab
Techniques: Deep Neural Networks, GridSearchCV, class imbalance handling (SMOTE), model blending, SHAP for interpretability
Output: Final DNN + CatBoost blended model, SHAP analysis, and full codebase with documentation published on GitHub
View on GitHub
Skills: Python · Machine Learning · Fraud Detection · Deep Learning · Data Science -
Predicting Healthcare Access Disparities Using Public Health Data
-
Designed and developed a machine learning pipeline to analyze and predict disparities in healthcare access across demographic groups using 2023 CDC NHIS data. The project involved exploratory data analysis, feature engineering, and training multiple models, including decision-tree based ensemble methods (Random Forest and LightGBM) to assess the impact of social, demographic, and health-related factors.
Key outcomes included identifying features most correlated with limited access and…Designed and developed a machine learning pipeline to analyze and predict disparities in healthcare access across demographic groups using 2023 CDC NHIS data. The project involved exploratory data analysis, feature engineering, and training multiple models, including decision-tree based ensemble methods (Random Forest and LightGBM) to assess the impact of social, demographic, and health-related factors.
Key outcomes included identifying features most correlated with limited access and addressing data bias due to racial imbalance in the source dataset.
Tools: Python, scikit-learn, LightGBM, Git LFS, Jupyter Lab
Techniques: Regression modeling, cross-validation, model interpretability
Output: Trained ML model and detailed report published on GitHub
View on GitHub
Honors & Awards
-
Great People, Great Performance Award
Microsoft
Languages
-
French
Full professional proficiency
-
Amharic
Native or bilingual proficiency
-
English
Native or bilingual proficiency
Recommendations received
3 people have recommended Biniam (Bini)
Join now to viewOther similar profiles
Explore more posts
-
Adnan Masood, PhD.
Just published in The AI Journal: we’re watching AI shift from clever copilots to accountable enterprise infrastructure. In 2026, it won’t be “what can the model do?” — it’ll be “what did the system deliver, at what risk, and at what cost?” I break down the three forces driving that transition (agents, compute, and a tougher economic/regulatory bar). Read More: https://lnkd.in/dPqQg4Ba #AI #EnterpriseAI #ResponsibleAI #DigitalTransformation
2 Comments -
Databricks
🏆 At the Databricks Grounded Reasoning Cup, academic teams from Stanford, UMass, Yale and other universities tackled a previously unseen corpus and set of tasks to answer one of the hardest questions in AI evaluation: how well do performance improvements on a benchmark generalize to similar, real-world tasks? Stanford’s winning team achieved 63.3% accuracy through an end-to-end agent optimization strategy combining a library of reusable skills, targeted document-representation fallbacks, and adaptive verification. See the techniques that set apart the winning teams: https://lnkd.in/gqkzHFVG
5 Comments -
Roberto Hortal
""Agent-native Architectures" discusses how LLMs operating in a loop with tools can achieve complex outcomes far beyond coding. This concept is fundamental to creating highly adaptive products. I find the focus on atomic tools & emergent capabilities particularly insightful! 👇 https://buff.ly/ErqjDEp #ProductManagement #AI 🤖
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content