Applied scientist and AI systems engineer with a Ph.D. from UIUC. Currently Senior AI Engineer at Doctronic, engineering closed-loop self-learning feedback engines (`#md-ai-feedback`), the NoHarm / MAST clinical safety programme (lifting F1 to 0.901, surpassing published SOTA leaderboards), and MedKG probabilistic factor graphs with NumPyro. Founder of UtterX and former Applied Scientist at AWS AI Labs co-leading enterprise customization for Amazon CodeWhisperer from inception to GA.
Closed-loop self-learning feedback engines, MedKG factor graphs, clinical safety programmes, autonomous agentic runtimes, and enterprise post-training recipes.
At Doctronic: Engineered Sol parity on Tier 2 clinical reasoning models matching Opus-high accuracy (94.0%), architected graph-grounded LLM judges against First Databank (FDB), and built 5-fold CV ablation harnesses.
Built an automated continuous learning loop ingesting live physician critiques from #md-ai-feedback, classifying failure taxonomies with Claude Opus 5.5, and generating automated regression suites and MIPRO prompt tuning.
Architected MedKG: a probabilistic clinical factor graph compiling into NumPyro for Bayesian inference and Value of Information (VOI) calculations, alongside in-process graph-grounded LLM judges (ADR-034).
Led the NoHarm / MAST clinical safety evaluation programme across 330 inputs, lifting benchmark F1 from 0.881 to 0.901 (surpassing published SOTA leaderboards), cutting harmful verdicts by -0.058, and lifting guideline grounding to 49%.
Founder & CEO project: Built an end-to-end self-healing agentic engine featuring programmatic verifiers, visual diff gates, and multi-turn verbal reinforcement loops.
Custom GPU kernel development computing unmerged low-rank adapter paths over an int8 quantized base model, fusing quantized GEMM with low-rank FP16 projections to cut multi-tenant serving costs.
At AWS AI Labs: Co-led applied research for CodeWhisperer enterprise customization from inception to GA, designing RAG + LoRAX multi-tenant adapter adaptation serving millions of developers.
Google Maps Geocoding GUI
Random Forest Regression to Impute Missing Values
Worth the Cost? An Examination of the Commercial Revitalization & Commercial Expansion Programs
Visualizing State/County Migration
Deferring maintenance expenditures in fiscal downturns can be more costly than accumulating debt or cutting pension contributions. This paper takes advantage of the formulary distribution of Federal highway grants to examine the extent and context of highway maintenance deferral. Since the formula factors used in the apportionment of Federal highway grants are always three years old, I can measure exogenous grant shocks as the difference between this year’s and previous year’s forecast of each state’s future grants. I show that state governments respond to a negative grant shock by cutting highway maintenance expenditures more rapidly than other expenditures. The Impulse Response Functions (IRFs) of maintenance expenditures as a share of total expenditures show that states only partially compensate for the initial cuts in the subsequent periods – deferring maintenance expenditures and accumulating maintenance needs. Furthermore, I find that the deferral of maintenance expenditures is more pronounced in election years, which suggests that this inefficient behavior is subject to agency problems.
Productivity and Quality-of-Life Benefits to Rural Infrastructure
Valuing Public Goods More Generally: The Case of Infrastructure
In-depth writeups on post-training, causal econometrics, cloud architectures, and clean code standards.
Engineering a 2×3 ft 12-hexagon illuminated wall relief on a Bambu P2S: solving the floating-island dilemma with a 0.8 mm translucent cap, side-loaded N52 magnet tunnels, and M3 heat-set shear pins.
Back in October, I got an email from Heroku saying that mLab MongoDB add-on is going away and I need to remove it from my Heroku App i.e. my personal website. You can use the free Atlas MongoDB to keep your KeystoneJS or other Heroku apps up and running.
"Even bad code can function. But if code isn't clean, it can bring a development organization to its knees. Every year, countless hours and significant resources are lost because of poorly written code. But it doesn't have to be that way." (Source: [Clean Code]) Here's how to fix that!
Free SSL with Free Heroku Dynos: Cloudflare to rescue
Get notified when new research on clinical reasoning, agentic systems, and post-training benchmarks are published. Syndicated across Substack, Beehiiv, and RSS.