Where a metric needs hardware I do not have, it is marked pending rather than claimed. Where a repository is private client work, it says so and shows no link. Filter by domain or search by name below.
Twelve of the repositories below are one platform, not twelve unrelated demos.
A mission-operations platform for a fictional satellite constellation, built as a set of standalone
engineering projects that fit together. Telemetry flows up from the data plane
(constellation, constellation-stream) through the agentic copilot
(groundstation) to the operator console (groundstation-console). The
copilot's reasoning core is trained by groundstation-train and grounded by the
retrieval systems (groundstation-rag, argus). Every command the copilot
proposes is authenticated by aegis and emitted only after a human approves it. The
whole thing is deployed by liftoff.
The reason it is worth building as a system rather than as isolated projects: the approval gate, the citation format, and the evaluation contract have to mean the same thing in every component, and that constraint is where the interesting engineering lives.
SFT then DPO on preference pairs where the chosen response defers to a human approval gate and the rejected one bypasses it. The eval harness is calibrated by two reference policies and the calibration is asserted in CI, so a broken grader fails loudly.
oracle 1.0 / bypass 0.0 on a 77-scenario held-out set · model lift pending GPU
A LangGraph multi-agent graph operating a live system through MCP tools and retrieval-cited playbooks, with tiered model routing by task cost. It halts at an approval gate and emits nothing until a human decides.
anomaly-type classifier: accuracy 0.967, macro-F1 0.931 on held-out satellites
Fifteen documented attack cases across five classes, each with a machine-checkable success oracle. Every defense toggles independently with per-defense attribution, so contributions are ablated rather than assumed. Runs offline and deterministically with no API key, so scores are CI-safe.
15 cases · 5 attack classes · quality gate fails the build below threshold
Decomposes a document into atomic, self-contained factual claims, verifies each against the live web in parallel, and returns a verdict plus sources per claim. The point is the schema: an annotation unit small enough that a reviewer can actually adjudicate it.
supported / contradicted / unverified per claim, with confidence and sources
Retrieval over a large engineering and standards corpus, answering with doc#section citations. Retrieval quality is a build gate: a regression in citation accuracy, recall, or ranking below threshold fails CI rather than quietly degrading.
citation-acc 0.967 · recall@k 1.0 · MRR 0.976
A deployed real-versus-synthetic image classifier behind a live API, built for a consumer-protection product. Thresholds set against the asymmetric cost of a false positive versus a false negative, with an eval set built to detect drift as generative models moved. Ships an honest model card.
0.996 accuracy · 0.9999 ROC-AUC on held-out test
Natural-language exploration of public earth-observation imagery. An LLM agent plans and calls geospatial tools, grounds answers with retrieval, and animates results onto a WebGL map. The same tools are exposed over MCP, so any MCP host can drive them. Runs offline with a deterministic scripted agent when no key is present.
public Sentinel-2 L2A only · deterministic offline mode
A compact UNet trained from scratch, no pretrained weights and no model download, segmenting surface defects across four classes. Exported to ONNX and wired back into argus as an enrichment step. Reproducible from a clean clone with one command.
mean IoU 0.882 on 160 held-out frames · per-class 0.76 to 0.95
Kafka to Flink to Iceberg, with stateful event-time windowing, bounded-out-of-orderness watermarks, allowed lateness that updates a fired window, and an exactly-once sink. Records that arrive too late are dead-lettered, never dropped. Two runtimes with identical semantics: a real Flink cluster and a cluster-free reference engine for CI.
22,388 rec/s · p99 query 1.31 ms at 1,000,000 rows
Telemetry data plane: streaming ingestion from Redpanda into TimescaleDB, unsupervised anomaly detection, a query and metering API, and a React ops console.
anomaly detection F1 0.720 (recall 0.808) over 180,000+ streamed records
Command authority for a system where a wrong command is expensive: Ed25519 signed envelopes chaining to an offline root, key rotation, replay protection, mutual TLS, and an adversarial harness with a measured, toggleable defense stack.
11/11 attacks land with defenses off · 0/11 with defenses on
A Java 11 port of the aegis cryptographic core, including a pure-Java Ed25519 implementation, proven correct by replaying the original's behavioural test vectors and reproducing every output byte for byte. Builds and tests with a plain javac harness and no third-party jars.
38 shared test vectors pass byte-for-byte across Python and Java
One declarative workload spec (the running example is a multi-GPU inference server) rendered for managed cloud, on-prem Kubernetes with a controller-runtime operator, and an air-gapped bare-metal bundle. Every artifact is validated by the real tool, not merely produced. GPU specifics are wired correctly in each target.
3 targets · byte-identical re-render · validated on amd64, arm64, darwin
The suite-level deploy layer: one spec rendered and validated for cloud (Terraform, EKS with GPU nodes), on-prem (Helm, CRD, operator), and air-gapped bare metal, with mTLS and Ed25519 secrets and SLO-driven rollback.
all 3 targets pass terraform, helm/kubeconform, and shellcheck
The operator console: a Go service owning telemetry and incident state behind an HTTP/JSON API fed live off Kafka, with an Angular SPA over it. Runs from a clean clone against a bundled deterministic fixture and switches to live telemetry when a broker is configured. Profiled with committed before-and-after artifacts.
fleet endpoint ~1,720x faster after profiling · SPA bundle -17.6%
A 3-DOF attitude-control simulator: real rotational dynamics via Euler's equations rather than a small-angle shortcut, a quaternion-PD controller that respects a finite actuator, and a fixed-rate loop that produces a bit-identical trajectory every run. Builds and tests with nothing but a C++20 compiler.
30° slew settles in 7.246 s to 0.0008° final error at 1 kHz
A unified real-time risk and analytics layer for event contracts held across two venues. Ingests positions and live market data, continuously recomputes one risk picture, and flags where the same real-world event is priced differently. Read-only and analytical: it places no orders.
boots on simulated data with zero secrets via docker compose
A per-position roster-value model comparing on-field production to contract cost, retraining itself daily via GitHub Actions and publishing to a live dashboard. The interesting part is data discipline: missing contract values are never imputed to zero, every player carries a match-quality flag, and low-sample players are routed to a committed diagnostic file rather than silently dropped.
retrains daily at 11:00 UTC · excluded players committed for audit
A paper portfolio (no broker, no real money) that marks to market each weekday via a GitHub Actions cron, computes benchmark-relative analytics and a data-quality check, updates a self-supervised per-ticker conviction bias, and writes a journal entry every run.
daily cron · live P&L table rewritten in the README each run
A small quantitative library on a pure-math foundation: Monte Carlo and Black-Scholes option pricing, Greeks, mean-variance portfolio optimization, and VaR/CVaR risk analysis.
Animated math explainers built on Manim Community, reproducing the visual grammar of 3Blue1Brown: equations animated in lockstep with the geometry they describe. The main scene builds a square wave from chained rotating vectors, with odd harmonics falling off like 1/n.
committed demo render · pinned dependency lockfile
The map for the twelve-repository system: how the pieces connect, the live headline metric for each, and the exact command that regenerates every number. A collector reads the committed metrics artifacts and reprints them, so the map cannot drift silently from the repositories it describes.
8 headline metrics, each backed by a committed artifact
This site. Static, dependency-free, hosted on GitHub Pages: hand-written CSS with a token layer, one shared script, an animated SVG diagram, and a command palette. No framework and no build step.
reduced-motion path for every animation on the site
The production system behind the correction loop on the home page. Rails and Postgres API, a client-facing JavaScript portal with OAuth 2.0 + PKCE and role-scoped authorization, a Snowflake reporting layer, automated build and deploy, render-escaping invariants, contract tests, and a full security-audit remediation round.
private repository · described, not linked
A LangGraph.js and Claude API agent platform in daily use by non-technical staff, shipped alongside the evaluation framework that measures whether its output is actually correct, including adversarial testing against a curated attack corpus.
private repository · described, not linked