Martin Mashalov

Energy Forecast Lab

Forecast the peak. Price the uncertainty. An auditable research system for Texas electricity demand and execution-aware paper trading.

Residual regression, whole-day error scenarios and chronological calibration turn an operator forecast into a distribution for the daily peak. Explore a full year of forecasts, all decision clocks, matched calibration controls and the liquidity screen that found no guaranteed package surplus.

The initial ridge model reduces peak error by 12.37% versus the operator, and 4.04% versus simple bias correction on 357 paired development days. The public extraction has 714 offline tests. Better forecasting is measured; trading profitability remains unproven.

WeatherPred

A weather signal has to survive the trade. A quantitative research system that follows it through fees, latency, queue position and settlement.

Original NOAA forecasts and NWS report versions, coherent probability models, and a resumable search over momentum, reversals, favorites, longshots and observed-temperature constraints. The paper ledger accounts for partial maker and taker fills, cash reservations and correlated exposure. An independent audit reconstructs 253,227 hypothetical trade calculations across alternative policies from raw market records.

The best-looking backtest is not the conclusion: the monthly strategy selector using only earlier results loses $4.45 on its conditional $100 account. Profitability remains unproven. Every strategy, equation, failed experiment and execution assumption is documented.

gpt2-harness

A training harness built from the parts up: all five parallelism strategies, a roofline, and a tool that tells you why a run is slow.

Data, sharded-data (ZeRO 1 through 3), tensor, pipeline and context parallelism, each written against torch.distributed and each checked against a single-process reference. The per-step wire volume is counted byte for byte and matched to its closed form, so ZeRO-2 and DDP are shown to move identical bytes, which is what you expect once you notice an all-reduce is a reduce-scatter followed by an all-gather. The pipeline bubble is measured against the analytic (p−1)/(m+p−1) for one to sixteen micro-batches. On top of that: a measured roofline, model-FLOPs utilisation, checkpoints that reshard across parallel layouts, a killed rank restarted from its last checkpoint, a resumable streaming loader, and Slurm, Ray and Kubernetes launchers.

The correctness proofs run on gloo because proving a sharded implementation matches a single-process reference does not need a GPU. Throughput and interconnect numbers do, so those are modelled from published bandwidths and labelled as modelled throughout. My thesis training is the other half of this: bfloat16 and FlashAttention-2 on H100s, with a subset of the evaluations on A100 and B200. Underneath it all, the GPT-2 implementation is verified to a worst logit disagreement of 6.1e−5 against HuggingFace across 75 layers, which is what makes the rest of the measurements mean something.

Sen  papyra.org

Describe a back-office process in plain English and get a workflow you can see, edit and run. Two insurance agencies use it. A person approves anything that leaves the building.

The description compiles into a visual node graph of triggers, tool steps, branches and approval gates, and a durable engine runs it per account across email, cloud storage, and the agency management system, through a registry of 64 tools. A workflow that sends real client mail must never send twice and never silently skip a step. Every state transition passes one chokepoint enforcing four invariants. Each one is traced to the production failure that caused it.

Reinforcement learning on diffusion language models

A diffusion language model chooses which position to reveal next. I tested whether post-training should change its weights, its decoding order, or both.

Diffusion language models unmask a sequence instead of writing left to right, so they face a choice autoregressive models never face: which position to reveal next. I ran a grid over three diffusion models, two post-training levers, and two reward designs on HumanEval and MATH-500, with MBPP and GSM8K held out. Carrying a learned ordering policy across to reward-tuned weights unchanged lowered accuracy; re-fitting it on those weights scored highest, though a continuation control means the size of that gain is not attributable to re-adaptation alone. The policy is fitted to a model, not to a task. And a checkpoint that had already been RL-tuned resisted every method: its within-group reward dispersion had collapsed, leaving GRPO almost nothing to learn from.

GridPull

Commercial insurance prospects built from public filings. Every renewal date is either quoted verbatim from a named document or labeled an estimate.

A producer needs to call a business ninety days before its policy renews. A verified renewal date is copied straight from the source registry and marked verified; where none exists the lead carries an estimate, marked as an estimate, and the only date arithmetic anywhere is “is it today or later”. A 150-worker asyncio pool streams document extraction over server-sent events, and the cross-sell engine is deliberately deterministic, with no model in the loop, because a coverage-gap recommendation has to be explainable.

gtm-agent

Finds people describing your problem on Reddit, scores them against an ideal-customer profile, and drafts the outreach. Nothing sends without a human approving it.

Cheap deterministic rules run first; the model only ever sees what survives them, and a test enforces that. Every model call is metered and attributed, so cost per qualified lead is measured end to end on a recorded corpus against a deterministic offline model. Official APIs only, no scraping and no browser automation of LinkedIn. That is the difference between a tool a company can deploy and one that gets its accounts banned.

agent-fleet

An agent run spends almost all its time waiting on a model API, so one asyncio supervisor handles thousands. Killing the process mid-run loses none of them.

An agent run is roughly 99% I/O wait on a model API, so the right shape is one asyncio supervisor with a bounded pool rather than a thread per agent. The hard parts come after: an at-least-once queue with leases and idempotency keys, per-step checkpointing so a killed process resumes mid-run, and backpressure that never lets concurrency exceed its cap. Measured in a 1-vCPU, 1 GB container against a deterministic mock provider: 546 agent steps a second.

generative-eval

"My generative model produces good samples" is not one claim. It is four, and they come apart.

Four questions, asked in order: does the metric work at all, does the model match the published marginals, can a classifier tell its samples from real ones, and is the synthetic data any use downstream. Seven families climb the same ladder under one budget: five GAN objectives, denoising diffusion and flow matching, read against a classical GARCH fit and two nulls nothing should lose to. Two testbeds where the answer is known independently: mixtures whose modes are exact, and seventeen years of real intraday equity bars whose stylized facts are a published answer key.

Flow matching is the first family here to beat the classical baseline on real data, and it holds all eight stylized facts down to ten sampling steps. Which families you benchmark decides what your benchmark concludes: over the five GANs alone, ranking by 2D mode coverage anti-correlates with ranking by detectability at Spearman −0.90.