AI Lead & Software Engineer Goldsmith & Co. New York

Systems that keep a human in the loop. And the evaluation that says whether they work.

I am Maisy Mylod. I am the sole engineer at an executive search firm, where I own every production system end to end: a Rails and Postgres API, a client-facing portal, a Snowflake reporting layer, and an internal agent platform in daily use by non-technical staff.

The part I care most about is the part that usually gets skipped. Not the model. The correction loop where real annotators fix the data, and the grader you calibrate before you let it tell you that you are making progress.

Selected work Experience GitHub ↗

Open to AI engineering roles

The correction loop goldsmith & co · production

human decision automated accepted corrections feed the system of record

Approach

Three commitments that show up in every repository below.

The interesting problem is rarely the model. It is the loop around it.

The gate is a design property

Not a review step bolted on afterward. In production, recruiters correct records inline, the correction lands in an override store, an admin is pinged to review, and only the accepted version reaches the system of record. Real annotators who are not engineers, on data the business depends on.

goldsmith & co · in production

Grade the grader first

An eval you have not calibrated is a number you cannot use. Two reference policies pin the scale: a gate-compliant oracle must score 1.0, a bypass baseline must score 0.0, and CI fails if that calibration drifts. A broken grader fails loudly instead of quietly reporting progress.

groundstation-train · enforced in CI

Every number regenerates

Nothing here is asserted. Each metric on this page comes from a held-out eval or benchmark committed in its repository, alongside the exact command that reproduces it from a clean checkout. Where a number needs hardware I do not have, it is labelled as pending, not claimed.

heliosnet · make eval

Measured

Held-out evals, committed in the repository named beside each row.

Measured, not asserted. Each row regenerates with one command.

Retrieval citation accuracy, top-1, on held-out QA with gold citations groundstation-rag · make eval · CI gate 0.90 0.967

Real-vs-synthetic image classifier accuracy, held-out test set athena-ai · EfficientNet-B0 · ROC-AUC 0.9999 0.996

Anomaly-type classifier accuracy, held-out satellites groundstation · make train · macro-F1 0.931 0.967

Defect-segmentation mean IoU, held-out frames, trained from scratch constellation-vision · make eval · per-class 0.76 to 0.95 0.882

Adversarial attacks blocked with the defense stack enabled aegis · make demo · 11/11 land with defenses off 11 / 11

Approval-gate eval calibration: oracle policy versus bypass baseline groundstation-train · make eval-refs · 77-scenario held-out set 1.0 / 0.0

One caveat stated up front, because it belongs up front: groundstation-train's model headline (a base versus SFT+DPO lift) needs a GPU fine-tune I have not run. What is committed and CI-enforced there is the eval harness itself. The training pipeline is real; the model number is honestly pending.

Selected work

Twenty-seven public repositories. These eight carry the thesis.

Built, evaluated, and shipped. Alignment, agents, adversarial testing, annotation, data, infrastructure.

Preference
alignment groundstation-train SFT followed by DPO on preference pairs where the chosen response defers to a human approval gate and the rejected one bypasses it. The eval harness is graded before it is trusted: an oracle policy pins 1.0, a bypass baseline pins 0.0, and the calibration is asserted in CI. oracle 1.0 · bypass 0.0 model lift pending GPU pythondpolora Agent
orchestration groundstation 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. accuracy 0.967 · macro-F1 0.931 langgraphmcprag Adversarial
evaluation gauntlet Fifteen documented attack cases across five classes (direct and indirect injection, jailbreak, exfiltration, tool abuse), each with a machine-checkable success oracle. Every defense toggles independently with per-defense attribution, so contributions are ablated rather than assumed. The whole evaluation runs offline and deterministically, with no API key. 15 cases · 5 attack classes pythonred-teamdeterministic Annotation
schema truth-editor Decomposes a document into atomic, self-contained factual claims, then verifies each one independently 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. per-claim verdicts + sources claude apiweb searchpython Retrieval
quality groundstation-rag 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 llamaindexqdrant Applied
modeling athena-ai A deployed real-versus-synthetic image classifier behind a live API, built for a consumer-protection product. Decision thresholds were set against the asymmetric cost of a false positive versus a false negative rather than optimizing a symmetric metric, with an eval set built to detect drift as the underlying generative models moved. 0.996 accuracy · 0.9999 ROC-AUC pytorchefficientnet-b0model card Data
infrastructure constellation-stream Kafka to Flink to Iceberg, with stateful event-time windowing, watermarks, allowed lateness, and an exactly-once sink. Records that arrive too late are dead-lettered, never dropped. 22,388 rec/s · p99 1.31 ms at 1M rows kafkaflinkiceberg Deployment
tooling outpost One declarative workload spec, rendered for managed cloud, on-prem Kubernetes, and an air-gapped bare-metal bundle. Every rendered artifact is validated by the real tool (terraform, helm, kubeconform, shellcheck), not merely produced. Output is byte-identical across re-renders. 3 targets · deterministic output goterraformhelm

Also public: aegis and aegis-java (Ed25519 command authentication, with 38 shared test vectors passing byte-for-byte across both languages), constellation, constellation-vision, groundstation-console, liftoff, argus, slew, prediction-market-platform, nfl-moneyball, apex-portfolio, quantlib-mm, and manim-explainer.

All projects GitHub ↗

Experience

Sole engineer, founder, and analyst. Full history on the experience page.

Production systems, owned end to end. Small teams, real users, no one else to hand it to.

Goldsmith & Co. AI Lead & Software Engineer

Jan 2025 — present · New York

Sole engineer: I own the Rails and Postgres API, a client-facing portal (OAuth 2.0 with PKCE, role-scoped authorization), a Snowflake reporting layer, and an internal AI platform, and I set technical direction with firm leadership.
Built the human correction loop for a production data pipeline, diagrammed above: non-engineer annotators correct records inline, an admin reviews, and accepted corrections feed back into the system of record.
Designed and shipped an agent platform (LangGraph.js and the Claude API) now in daily use by non-technical staff, plus the evaluation framework measuring whether its output is correct, including adversarial testing against a curated attack corpus.
Diagnosed a silent 8% data loss in a production sync pipeline (4,840+ failed syncs, missing history rows, phantom records) that had corrupted downstream reporting undetected. Traced root cause, repaired the historical record, and added invariant checks so the failure now surfaces immediately.
Supervised an ML intern on a learning-to-rank project (LambdaMART, NDCG). Established by measurement that ranking quality was capped by label sparsity rather than model choice, and redirected the work instead of tuning against a ceiling.

railspostgreslanggraphclaude apisnowflakeoauth 2.0 / pkce

Athena Founder & Engineer

2025 — present · New York

Founded a consumer-protection startup detecting synthetic media and non-consensual imagery. Applied to Y Combinator.
Trained and deployed an image classifier (EfficientNet-B0, PyTorch) with perceptual hashing for near-duplicate matching: 0.996 accuracy on a held-out set of real and synthetic images.
Set decision thresholds against the asymmetric cost of a false positive versus a false negative, and maintained an eval set built to detect drift as the underlying generative models moved.

pytorchefficientnetperceptual hashingeval design

CLEAR Data Analyst

Jan 2024 — Jan 2025 · New York

Built churn forecasting models across a 20M+ member base at a biometric identity company, and ran scenario analysis on how pricing changes traded retained volume against revenue.
Automated a $1B+ revenue forecast, replacing a hand-assembled analysis with a reproducible pipeline that could be re-run and checked against realized results.

snowflakesqlpythonforecasting

Full work history

Output

GitHub, 2026 year to date. Read from the contributions API on 4 August 2026.

I write a lot of code. Most of it is public and reproducible.

contributions this year

commits pushed

pull requests opened

repositories contributed to

The labels were the bottleneck, not the model.

The conclusion from the learning-to-rank work above, and the reason this page leads with annotation and evaluation rather than with architecture.

Get in touch

I would like to work on this at a larger scale. If that is the problem you are working on, let's talk.

Based in New York. Reachable at maisymylod@gmail.com.

Contact GitHub ↗ LinkedIn ↗

gh repo list maisymylod --limit 30