104 things I've grown, from clinical LLM systems to causal studies to weekend ML experiments. New projects sprout here on their own as I push them to GitHub. The big blooms are below; pick a patch to wander through the rest.
semantic search · your phrase and every project are embedded with OpenAI, then ranked by cosine similarity ✦
An intruder-detection system fusing simulated Wokwi hardware sensors with computer vision (YOLO person + weapon detection, fire, and face recognition), wired to a Streamlit dashboard with real-time email alerts.
A multimodal medical-record companion unifying RAG, document understanding, speech, and vision. Consensus extraction across LLMs hit 85.1% micro-F1 with sub-2s latency.
Benchmarked KIVI quantization, TopK sparsity, SnapKV eviction & MLA on Llama-2-7B with Triton kernels, 4× cache compression, 1.93× faster decode, 3.1× peak throughput.
End-to-end causal inference study of the Moertel (1990) colon cancer trial (n=929), evaluating average and heterogeneous treatment effects, mediation pathways, confounding bias, and external validity through transportability analyses.
A multi-agent CrewAI system for U.S. federal legal analysis: semantic USC retrieval, precedent search, elements analysis, and draft generation.
AI food app for video-to-recipe extraction, pantry matching, and personalized health coaching. React + Vite + Framer Motion on Netlify.
every project embedded and projected to 2D, so similar work sits close together. drop in something you care about and see where you land.
mapping the galaxy… 🌌
A multimodal medical-record companion unifying RAG, document understanding, speech, and vision. Consensus extraction across LLMs hit 85.1% micro-F1 with sub-2s latency.
A multi-agent CrewAI system for U.S. federal legal analysis: semantic USC retrieval, precedent search, elements analysis, and draft generation.
Full-stack RAG pipeline turning natural language into SQL and visualizations for sports performance analytics.
AI food app for video-to-recipe extraction, pantry matching, and personalized health coaching. React + Vite + Framer Motion on Netlify.
Converts cooking videos into structured recipes via a multi-stage vision-language pipeline: frame extraction, visual understanding, LLM reasoning.
Student-built dashboard for Columbia MSDS course reviews, live Google-Sheets data, AI-summarized reviews, rankings, and side-by-side comparisons.
AI coaching that maps fitness queries to personalized exercises via two-stage retrieval + LLM re-ranking. FastAPI, PostgreSQL, Claude.
AI diary turning journal entries into personalized verses + music recs. DistilRoBERTa emotion, K-Means over 867 songs, FAISS RAG, GPT-4o-mini.
AI blog-post generator with DALL·E images, SEO-optimized content with customizable tone, length, and generated visuals.
Educational medical-image analysis with Gemini Vision, upload or camera capture, safe non-diagnostic insights with built-in disclaimers.
Explains technical concepts through personalized analogies based on your interests. Interactive and friendly.
The accuracy-vs-cost Pareto frontier for local LLM inference on Apple silicon, energy included
Fault-injection harness for structured LLM outputs: JSON mode vs function calling vs grammar constraints vs a staged repair pipeline.
Adaptive reasoning-depth controller: early-exit chain-of-thought when the answer stabilizes, vs fixed budgets and self-consistency.
Regression testing for LLM reasoning: declarative test specs, consistency checks, and a CI gate that fails deploys on category regressions.
A false-positive rate for LLM safety behaviour: does the model refuse legitimate, sensitive-adjacent requests?
Does agreement across reasoning chains predict correctness? Calibration curves, ECE, and five selection strategies vs chain count.
Defensive benchmark: do agent architectures reduce prompt-injection susceptibility? Inert canary tokens only
Long-context stress tests: lost-in-the-middle heatmaps, multi-hop joins, instruction retention, distractors, and the latency and KV-memory bill.
Cost-, latency-, and quality-aware LLM model router benchmarked against always-large and always-small baselines over a 4k-request workload.
A complaint queue for bank complaints teams: ranks CFPB complaints by how likely each is to end with money returned, explains every score, and writes an analyst brief with Claude.
Live cross-model benchmark: reasoning accuracy on known-answer problems, real streaming latency and throughput, and structured-output validity across GPT-4o, Claude Sonnet 5, Gemini 2.5 Flash and Grok-4.3. Bring your own keys.
Production-style LLM gateway in pure Python stdlib: auth, semantic cache, rate limiting, retries, fallbacks, circuit breaker, cost tracking. Load-tested through a brownout.
Speculative decoding benchmark: draft size x gamma x task sweep with closed-form speedup validated by token-level Monte Carlo, plus a real Colab notebook.
LLM inference profiler: TTFT, inter-token latency, throughput, memory and quality across FP16/INT8/INT4, batch size, context length, KV caching. Dashboard included.
Delta debugging (ddmin) for LLM prompts: shrink a failing prompt to the minimal sub-prompt that still reproduces the failure.
Defensive LLM safety evaluation: refusal robustness across harm categories and framings, over-refusal on benign lookalikes, consistency. No attack content.
A multi-agent CrewAI system for U.S. federal legal analysis: semantic USC retrieval, precedent search, elements analysis, and draft generation.
IEOR 4570 Project 1: a tool-calling agent deployed to Cloud Run
An agent memory layer benchmarked on facts that change: contradiction resolution and recency vs naive similarity
Compare agent architectures at matched cost, not matched accuracy: the cost/accuracy frontier
Record-and-replay behavioural diffs for agent changes: path divergence vs outcome equivalence
Defensive benchmark: do agent architectures reduce prompt-injection susceptibility? Inert canary tokens only
Debug multi-agent systems like distributed systems: happens-before graphs and deadlock/livelock/lost-update detection
Full-stack RAG pipeline turning natural language into SQL and visualizations for sports performance analytics.
Student-built dashboard for Columbia MSDS course reviews, live Google-Sheets data, AI-summarized reviews, rankings, and side-by-side comparisons.
AI diary turning journal entries into personalized verses + music recs. DistilRoBERTa emotion, K-Means over 867 songs, FAISS RAG, GPT-4o-mini.
TF-IDF + Linear SVM on the WELFake dataset with a real-time credibility-prediction interface.
A summariser where every sentence carries its source span and unsupported ones are flagged, with a validated NLI verifier
Measures what identical content costs in different languages across three tokenizer generations, using 377 real translations of the Universal Declaration of Human Rights as a parallel corpus.
Look up an NYC building's official housing record before signing a lease: violations, complaints, bedbugs and evictions from NYC Open Data, compared per apartment with similar buildings, summarized in plain English.
End-to-end causal inference study of the Moertel (1990) colon cancer trial (n=929), evaluating average and heterogeneous treatment effects, mediation pathways, confounding bias, and external validity through transportability analyses.
ML & predictive analytics framework for identifying high-risk child-welfare cases using NCANDS data.
Benchmarking causal discovery under realistic measurement error, with a reliability-correction fix
A mediation analysis where the reader turns the untestable assumptions: DML effects with an interactive sensitivity dial
End-to-end causal inference study of the Moertel (1990) colon cancer trial (n=929), evaluating average and heterogeneous treatment effects, mediation pathways, confounding bias, and external validity through transportability analyses.
Visual analysis of how diet and lifestyle contribute to colorectal cancer risk, built in R/Shiny.
Shape bucketing as an optimisation over a real request-length distribution: DP-optimal buckets vs the recompilation storm
ML & predictive analytics framework for identifying high-risk child-welfare cases using NCANDS data.
Predicts used-car prices (India), 65.6% R² with XGBoost + Optuna, SHAP explainability across 4 ensembles.
Gradient Boosting on physicochemical properties with SHAP, feature importance, and PDPs for interpretability.
Production-style pipeline + Gradient Boosting for loan approval with SHAP explanations and an interactive UI.
XGBoost regression on California housing with an interactive what-if explorer.
Gradient Boosting on clinical indicators with SHAP for global and per-patient explanations.
Classifies sonar signals as rock or mine via cross-validated automated model selection.
Predicts repeat-purchase behavior for food-delivery businesses and surfaces key retention drivers.
EDA, feature engineering, SMOTE for imbalance, and tuned models exploring heart-disease risk factors.
Dataset-free per-rep form feedback from your own best reps, validated on exactly-known degradations
A logging simulator with known ground truth that maps where off-policy evaluation estimators lie
A gallery of small environments where a plausible reward produces a degenerate policy, plus an honest look at detection
Offline RL where the headline is the confidence interval's width and the power analysis, not a fragile point estimate
Evaluates RL curricula on tasks harder than anything they covered, holding the task set and training budget fixed so only the ordering varies. MuJoCo cart-pole family, staged cross-entropy policy search.
Scores generated trajectories by whether the dynamics could have produced them, using the fact that a mechanism's one-step reachable set is a line segment. Untrained and parameter-free, compared against a learned discriminator.
Active learning for hand-eye calibration: D-optimal pose selection reaches target accuracy in far fewer poses
Decomposes the sim-to-real gap into per-mismatch line items, then tests whether the decomposition is additive. Five independently switchable reality mismatches on a MuJoCo cart-pole, with domain randomization as the follow-up.
Benchmarks how well a controller recovers from a disturbance rather than whether it succeeds from nominal states. MuJoCo cart-pole, LQR gains solved from the simulator's own linearization, recovery basins found by bisection.
Measures real perception latency and real OS scheduling jitter under process contention, then drives a MuJoCo plant with the measured schedules to see whether commodity timing is good enough to close a robot control loop.
Map where annotators disagree over preference space, and how that reshapes reward models
Ranked grasp hypotheses with stated reasons and an abstention option, plus the success-vs-coverage curve
Synthetic control on a tiny donor pool with honest placebo-permutation inference for small places
Auto-propose and run the negative-control falsification tests an observational analysis should have failed
What ordinary handling does to an image watermark: survivability cards for a block-DCT scheme (defensive)
Generate the conditions that break your detector, measure the per-edit taxonomy, then train on the failures
The smallest file that keeps the detector right: per-image JPEG bitrate policy vs a global quality
AutoML that refuses features on cost grounds, not just accuracy: joint accuracy-cost Pareto selection with a per-feature argument
Measures how many labelled examples a hand-written keyword rule substitutes for, swept across data budgets and rule counts. Real 20 Newsgroups, with rules frozen from a held-out seed split so they add knowledge rather than re-encode the training set.
A leakage audit for image datasets: masked-variant ablations that fingerprint what the dataset really teaches
Recover true labels and annotator reliabilities from noisy votes with Dawid-Skene, and verify only the suspicious labels
Cluster, name, and validate a classifier's failure modes, then prove the names identify held-out errors
An audit of what a distilled student actually kept: per-class deltas, agreement with the teacher, accuracy banded by the teacher's confidence, and calibration. Every student matched against a same-size control trained from scratch.
Measures the false-positive rate of small-seed comparisons by training identical configurations 40 times, with a provably null intervention as the control and resampled comparisons at every seed count.
Measures both costs of asking a validation set to choose when to stop and to report the result, across validation sizes from 100 to 2000, against a clean test set never used for any decision.
Does any data ordering beat shuffling at fixed compute, once scoring cost is counted? An honest negative
Catastrophic forgetting as a dose-response law, inverted into a calculator for the largest safe fine-tune
A post-mortem for failed training runs: reads telemetry, names the failure mechanism, prescribes a change, and retrains to verify the fix actually works. Real MNIST with injected faults and a held-out fault battery.
Rewrite text so it means the same but no longer sounds like you: the anonymity-meaning frontier
A model that knows when a question is underspecified: answer-entropy ambiguity detection with clarifying questions
Ranks real robot teleoperation episodes (DROID) by how likely each is a failed demonstration, from the arm's motion alone, so a reviewer watches the likely failures first.
A reusable harness and readiness report for deciding whether an intent router is safe to launch: in-scope accuracy, false accepts, wrongful declines and cost, at every setting.
A product experiment readout on a real randomized email test of 64,000 customers: soundness checks, effects with intervals, who each email works for, and a decision.
Measures what scanner-invariance costs on real pathology slides, comparing five preprocessing strategies across four sites. Domain shift built from H&E stain deconvolution rather than added noise.
Learn control from pairwise trajectory preferences with a direct objective, no reward model
ML & predictive analytics framework for identifying high-risk child-welfare cases using NCANDS data.
Predicts used-car prices (India), 65.6% R² with XGBoost + Optuna, SHAP explainability across 4 ensembles.
Production-style pipeline + Gradient Boosting for loan approval with SHAP explanations and an interactive UI.
XGBoost regression on California housing with an interactive what-if explorer.
Gradient Boosting on clinical indicators with SHAP for global and per-patient explanations.
Predicts repeat-purchase behavior for food-delivery businesses and surfaces key retention drivers.
EDA, feature engineering, SMOTE for imbalance, and tuned models exploring heart-disease risk factors.
Diachronic term drift on arXiv with a seed-noise null and a held-out predictive test
Forecast when a tabular model will decay from early drift dynamics, with calibrated conformal intervals
When does gradient boosting stop being the right model? Learning-curve crossovers across 20 tabular datasets, with a meta-feature recommender
Predict the grokking (delayed-generalisation) jump from the first 10% of training, via gradient-noise dynamics
A cost model that picks a training configuration without running any of them, then scored against an exhaustive sweep of all 48 measured end to end. Real transformer training across batch, sequence length and activation checkpointing.
Does agreement across reasoning chains predict correctness? Calibration curves, ECE, and five selection strategies vs chain count.
Automated keratoconus detection using SVM and deep neural networks.
Classifies CT scans into Normal/Cyst/Stone/Tumor via VGG19 & ResNet50 transfer learning, 99.2% accuracy.
Automated cataract detection using CNNs and transfer learning.
VGG16 transfer learning + fine-tuning on GTSRB, 98% accuracy across 43 classes.
ResNet50 transfer learning across 38 plant-disease classes with training, evaluation, and inference tools.
Deep neural networks combining static + dynamic analysis to detect malicious Android apps.
When does gradient boosting stop being the right model? Learning-curve crossovers across 20 tabular datasets, with a meta-feature recommender
A cost model that picks a training configuration without running any of them, then scored against an exhaustive sweep of all 48 measured end to end. Real transformer training across batch, sequence length and activation checkpointing.
An intruder-detection system fusing simulated Wokwi hardware sensors with computer vision (YOLO person + weapon detection, fire, and face recognition), wired to a Streamlit dashboard with real-time email alerts.
A multimodal medical-record companion unifying RAG, document understanding, speech, and vision. Consensus extraction across LLMs hit 85.1% micro-F1 with sub-2s latency.
Converts cooking videos into structured recipes via a multi-stage vision-language pipeline: frame extraction, visual understanding, LLM reasoning.
Educational medical-image analysis with Gemini Vision, upload or camera capture, safe non-diagnostic insights with built-in disclaimers.
Automated keratoconus detection using SVM and deep neural networks.
Classifies CT scans into Normal/Cyst/Stone/Tumor via VGG19 & ResNet50 transfer learning, 99.2% accuracy.
Automated cataract detection using CNNs and transfer learning.
VGG16 transfer learning + fine-tuning on GTSRB, 98% accuracy across 43 classes.
ResNet50 transfer learning across 38 plant-disease classes with training, evaluation, and inference tools.
A free touch sensor: audio auto-labels contact, a vision model learns it, then the microphone is thrown away
Benchmarked KIVI quantization, TopK sparsity, SnapKV eviction & MLA on Llama-2-7B with Triton kernels, 4× cache compression, 1.93× faster decode, 3.1× peak throughput.
Diagnose out-of-memory errors that are fragmentation, not capacity: a caching-allocator replay, a fragmentation-aware OOM detector, and a size-class fix that removes the OOM without adding memory
Search which layers actually need their bits: evolutionary per-layer mixed-precision quantization vs uniform
Places three numerically identical attention implementations on a roofline measured on the device that runs them. Naive, a Python flash-style loop, and the framework's fused kernel, from 256 to 8192 tokens.
Live cross-model benchmark: reasoning accuracy on known-answer problems, real streaming latency and throughput, and structured-output validity across GPT-4o, Claude Sonnet 5, Gemini 2.5 Flash and Grok-4.3. Bring your own keys.
LLM inference profiler: TTFT, inter-token latency, throughput, memory and quality across FP16/INT8/INT4, batch size, context length, KV caching. Dashboard included.
Deep neural networks combining static + dynamic analysis to detect malicious Android apps.
Secure storage of encrypted medical/review data using ECDSA signatures, Proof of Work, and hash chaining.
A modular confusion/diffusion pipeline with swappable primitives (perm, spectral, xor, latin, inn), and a metrics harness (entropy, correlation, NPCR, UACI).
An intruder-detection system fusing simulated Wokwi hardware sensors with computer vision (YOLO person + weapon detection, fire, and face recognition), wired to a Streamlit dashboard with real-time email alerts.
A free touch sensor: audio auto-labels contact, a vision model learns it, then the microphone is thrown away