01 / AI systems researchCourse research · 2026
Turn failure
into feedback.
Symbiotic-SWE
I built a neuro-symbolic software-repair system that combines LLM-generated patches, program slicing, solver checks, counterexample-guided feedback, and real tests. In a held-out SymPy study, the full loop resolved more tasks—at a higher token and runtime cost.
Explore the case study ↗ The held-out result
Neural only2 / 1414.3% resolved
Full feedback loop5 / 1435.7% resolved
One repository family; the stronger result cost more tokens and time.
02 / AI systems researchCourse research · 2026
Build the loop.
Test what survives.
Molecule Generation with Reinforcement Learning
I implemented an end-to-end GRPO-style post-training path for DiffDock, then tested whether an apparent improvement held across another seed. It did not—and that instability became the most useful finding.
Explore the case study ↗ The honest result
Seed 44ImprovedMean RMSD 7.49 → 3.03 Å
Seed 45UnstableSuccess@1 0.80 → 0.40
A functional training loop, not yet a robust benchmark result.
03 / Efficient inference researchFive-person course collaboration · 2026
Evict less useful
context.
Semantic-Aware Chunk-Level KV Cache Eviction
I helped build a semantic approach to evicting whole context chunks from an LLM cache. A small live warmup recovered cache capacity without changing answer quality at a moderate budget, while exposing latency and robustness costs that still need work.
Explore the case study ↗ The live warmup
Cache released713 blocksEstimated 654 MB
Tradeoff+56% latency0% answer-quality delta
50 MuSiQue examples at 30% eviction; early evidence, not a production claim.
04 / Creative instrumentsCompleted class project · 2024
One gesture.
Two ways to understand it.
Pure Data Synthesizer + Visualizer
A MIDI keyboard becomes a control surface for sound and visuals. Notes, knobs, and an arpeggiator connect musical expression with a changing visual scene.
Explore the case study ↗ The interaction
Your MIDI keyboardNotes + eight knobs
↓
Pure Data patch
↙ ↘
Hear itSynth + effects
See itGEM visuals
System overview, based on the original project documentation.
05 / Attention & everyday workflowsIn development
Daily Focus Coach
How can a tool help someone return to what matters?
Daily Focus Coach is my next product project after Portfolio v1. The homepage interaction illustrates the direction: helping someone return to what matters. It is not yet a product demo, trial result, or persistent personal-data system.
See the illustrative interaction ↗ 