Adam A. Holmes
Hi! I’m a Computational Physicist and AI Researcher (Ph.D. Theoretical Physics, Cornell). I build accurate and efficient computational systems by combining paradigms: deterministic with stochastic, symbolic with neural, exact with approximate.
This thread runs through all my work. In quantum chemistry, I combined deterministic wavefunction selection with stochastic perturbation theory to make exact calculations tractable at scale (SHCI, 2,300+ citations). Now I apply the same thinking to AI, pairing symbolic reasoning with learned neural components so that each handles what it’s best at, and pushing on the efficiency of LLM inference and training. I’ve been building production AI systems since 2018 (Transformer-based semantic search, pre-BERT), deployed my algorithms on some of the largest supercomputers in the world at Lawrence Livermore National Lab, and built quantitative models for systematic trading at Citadel.
Large Language Models
| Project | Description | Tech |
|---|---|---|
| Inference Engine + Post-hoc MLA | From-scratch single-GPU inference engine for Qwen3, and a study of post-hoc Multi-head Latent Attention (MLA) conversion: compressing the KV cache 4× for +16% perplexity against a matched-budget control, via activation-aware SVD with a partial-RoPE split. A small mergeable LoRA corrector trained to minimize total-variation distance to the original beats the standard KL objective on every fidelity measure (80% vs 74% token agreement at 4×); the same corrector lifts speculative-decoding throughput to 1.48–1.60×. Custom Triton kernels and a bit-exact validation harness against HuggingFace. | PyTorch Triton CUDA |
| NanoGPT Single-GPU Harness | A single-GPU (16 GB) adaptation of the modded-nanogpt speedrun for architecture screening — a research harness, not a benchmark, where the value is in the rankings and a methodology built not to fool itself: an every-weight-actually-trains assertion, a unigram-floor gate, paired same-seed comparisons, per-variant learning-rate matching, and budget-stability re-runs — all CI-enforced on every push. | PyTorch NorMuon Muon |
Neurosymbolic AI
| Project | Description | Tech |
|---|---|---|
| Neurosymbolic Chess Engine | Chess engine where symbolic reasoning (mate search, material-aware quiescence search) gates neural evaluation. Symbolic knowledge dramatically accelerates learning: +600 Elo over the pure neural baseline (AlphaZero-style) in 20 vs 30 generations of self-play. | Rust MCTS PyTorch |
| Geometry Theorem Prover | Neurosymbolic prover that combines exhaustive symbolic deduction (49 rules to fixed point) with neural-guided MCTS over auxiliary constructions. A 4M-parameter transformer learns the creative step that deduction alone can’t do. Solves 189/231 problems on AlphaGeometry’s JGEX benchmark, including Morley’s theorem and the 9-point circle. | Rust PyO3 PyTorch |
Quality-Diversity & Multi-Agent Systems
| Project | Description | Tech |
|---|---|---|
| MMR-Elites | Quality-Diversity algorithm that reformulates archive maintenance as submodular maximization via Maximum Marginal Relevance from information retrieval. Fixed O(K) memory, O(K log K) selection, 12x better uniformity than MAP-Elites in 20-dimensional behavior spaces. | Rust PyO3 Python |
| Multi-Agent Path Planning | Optimal multi-robot navigation in Rust: Conflict-Based Search (CBS) for globally optimal, collision-free paths, with Optimal Reciprocal Collision Avoidance (ORCA) for real-time local avoidance. Python bindings via PyO3; 176 tests covering the search and collision-geometry guarantees. | Rust PyO3 CBS ORCA |
Optimization & EDA
| Project | Description | Tech |
|---|---|---|
| Macro Placement | Pairs a differentiable global placer (GPU — optimizes any loss you can write down) with a discrete greedy detailer on the exact metric: the continuous + exact combination that runs through my work. 31% below a RePlAce-quality reference across all 17 IBM chip benchmarks, with a paired-experiment appendix of what didn’t work. [write-up] · [code] | PyTorch CUDA NumPy |
Quantum Chemistry
| Project | Description | Tech |
|---|---|---|
| RISQ | Rust implementation of Semistochastic Heat-Bath Configuration Interaction (SHCI) — the method I invented during my Ph.D. — for near-exact electronic structure calculations. Combines deterministic selection of important wavefunction components with stochastic perturbative corrections; bitstring determinant representation, Davidson eigensolver, and alias sampling for O(1) stochastic draws. | Rust |
Foundational Research
My Ph.D. research introduced Heat-Bath Configuration Interaction (HCI) (Holmes et al., JCTC 2016) and Semistochastic HCI (Sharma, Holmes et al., JCTC 2017), methods that replaced inefficient generate-and-test approaches with deterministic selection of the most significant wavefunction components, combined with stochastic sampling for perturbative corrections. This deterministic + stochastic combination made previously intractable calculations routine.
These methods enabled the first near-exact potential energy surfaces for fourteen electronic states of the carbon dimer (Holmes et al., JCP 2017) and the ground-state binding curve of the chromium dimer (Li, Yao, Holmes et al., Phys. Rev. Res. 2020), a grand-challenge problem that had remained outstanding for decades. SHCI is now a leading benchmark method implemented in major quantum chemistry packages.
Contact
I’m always happy to chat about research, projects, or opportunities. Reach me via email or on LinkedIn.