diff --git a/job_applications/2026-09-07/01_frontier_ai_lab_safety_evaluations.md b/job_applications/2026-09-07/01_frontier_ai_lab_safety_evaluations.md
new file mode 100644
index 0000000..a18d549
--- /dev/null
+++ b/job_applications/2026-09-07/01_frontier_ai_lab_safety_evaluations.md
@@ -0,0 +1,42 @@
+# Cover letter — Frontier AI lab, safety / model evaluations
+
+**Archetype:** Anthropic, OpenAI, DeepMind, Scale AI Frontier Risk,
+Redwood, METR, MATS/Constellation, AISI-style employers hiring for
+model evaluations, red-teaming, or alignment research engineering.
+
+---
+
+Dear {HIRING_MANAGER},
+
+I'm applying for the {ROLE_TITLE} role at {COMPANY}. I'm finishing an
+M.S. in Applied Statistics at RIT (2026) and my independent portfolio
+is built around the question your team lives with every day: how do you
+evaluate frontier-model behavior in a way that is measurable, reliable,
+and defensible under review?
+
+My **AI Safety Red-Team Evaluation** project designs a two-stage
+workflow that pairs an LLM ensemble with a supervised harm classifier
+across 12,500 response pairs and six harm categories. I report
+Krippendorff's α = 0.81 for annotation reliability and 96.8% held-out
+classification accuracy, and — this is the part I think matters most —
+I keep annotator agreement, classifier performance, and downstream
+Bayesian risk analysis as three separate claims so a reviewer can
+audit each one on its own. The **LLM Ensemble Textbook Bias Detection**
+project extends the same pattern to 4,500 passages and 67,500
+LLM-as-judge ratings, with publisher-level effects modeled as Bayesian
+partial pooling in PyMC so that the uncertainty on a small publisher's
+score is honest rather than optimistic.
+
+{ROLE_SPECIFIC_HOOK}
+
+I'd bring rigorous evaluation design, comfort with PyMC / MCMC
+diagnostics and SHAP, and a habit of naming a project's limits in the
+same document as its results. I'm US-authorized and open to remote or
+on-site in New York; happy to walk through either project end-to-end
+in a first conversation.
+
+Best,
+Derek Lankeaux
+[LinkedIn](https://linkedin.com/in/derek-lankeaux) ·
+[GitHub](https://github.com/dl1413) ·
+[Portfolio](https://dl1413.github.io/LLM-Portfolio/)
diff --git a/job_applications/2026-09-07/02_nyc_fintech_data_scientist.md b/job_applications/2026-09-07/02_nyc_fintech_data_scientist.md
new file mode 100644
index 0000000..519c3b6
--- /dev/null
+++ b/job_applications/2026-09-07/02_nyc_fintech_data_scientist.md
@@ -0,0 +1,44 @@
+# Cover letter — NYC fintech / quant data scientist
+
+**Archetype:** JPMorganChase, Goldman Sachs, Morgan Stanley, BlackRock,
+Two Sigma, Point72 Cubist, Bridgewater, Citadel, Jane Street data-side
+hires — analytics / research roles that want statistical rigor plus
+LLM literacy for the growing internal-AI workstreams.
+
+---
+
+Dear {HIRING_MANAGER},
+
+I'm writing about the {ROLE_TITLE} position at {COMPANY}. I'm an M.S.
+Applied Statistics candidate at RIT (2026) with an independent project
+portfolio that leans on the two things I understand your desk cares
+about: calibrated uncertainty on the statistical side, and disciplined
+LLM evaluation on the emerging-tools side.
+
+The **LLM Ensemble Textbook Bias Detection** project is the clearest
+example. I ran 4,500 passages through a rubric-based LLM-as-judge
+ensemble for 67,500 total ratings, reported Krippendorff's α = 0.84
+for inter-rater reliability, and then modeled publisher-level effects
+with Bayesian partial pooling in PyMC with full MCMC diagnostics. The
+same habits show up in the **AI Safety Red-Team Evaluation**: 12,500
+response pairs, α = 0.81, 96.8% held-out accuracy, with SHAP
+attributions attached to the risk analysis so that a reviewer can see
+exactly which features are driving a flagged decision. My **RAG
+Production Pipeline** work adds the systems side — hybrid retrieval,
+grounding checks, and 94.2% citation precision — which is directly
+relevant to any internal knowledge-assistant effort your team is
+running or planning.
+
+{ROLE_SPECIFIC_HOOK}
+
+I'm comfortable in Python, SQL, and R; fluent with scikit-learn,
+XGBoost, LightGBM, PyMC, and MLflow; and I write results the way a
+partner or a regulator would want to read them — with the limitations
+listed in the same paragraph as the headline number. Based in the
+region, open to on-site NYC or hybrid, US-authorized.
+
+Best,
+Derek Lankeaux
+[LinkedIn](https://linkedin.com/in/derek-lankeaux) ·
+[GitHub](https://github.com/dl1413) ·
+[Portfolio](https://dl1413.github.io/LLM-Portfolio/)
diff --git a/job_applications/2026-09-07/03_llm_evaluation_startup.md b/job_applications/2026-09-07/03_llm_evaluation_startup.md
new file mode 100644
index 0000000..adff79f
--- /dev/null
+++ b/job_applications/2026-09-07/03_llm_evaluation_startup.md
@@ -0,0 +1,42 @@
+# Cover letter — LLM evaluation / observability startup
+
+**Archetype:** Braintrust, Patronus AI, Arize AI, Weights & Biases,
+LangSmith / LangChain, HumanLoop, Galileo, Log10, Confident AI —
+companies whose product is helping other teams evaluate, monitor, and
+improve LLM applications.
+
+---
+
+Dear {HIRING_MANAGER},
+
+I'm applying for {ROLE_TITLE} at {COMPANY}. Your product exists because
+teams shipping LLM features can't rely on a single accuracy number to
+know whether the thing is working; my three primary portfolio projects
+are, honestly, exactly that argument told from the user's side.
+
+**AI Safety Red-Team Evaluation** is a full evaluation workflow across
+12,500 response pairs and six harm categories, reported with
+Krippendorff's α = 0.81 alongside 96.8% held-out classifier accuracy —
+kept as separate signals rather than collapsed into one metric.
+**LLM Ensemble Textbook Bias Detection** runs 4,500 passages through
+a rubric-based judge ensemble for 67,500 ratings and layers a Bayesian
+hierarchical model on top so publisher-level effects have honest
+uncertainty. **RAG Production Pipeline** connects retrieval quality
+(96.3% Recall@10), grounding (94.2% citation precision), and
+operational concerns — latency, drift, failure modes, rollback — in
+one system. If your customers are asking "what should I be measuring
+and how do I know it's stable," these are the questions I've been
+sitting with.
+
+{ROLE_SPECIFIC_HOOK}
+
+I'd bring an evaluations-first mindset, real work in PyMC and MLflow,
+enough FastAPI and Docker to prototype something usable, and a habit
+of writing docs the way I'd want a customer to read them. Remote is
+ideal for me; I'm US-authorized and can start conversations right away.
+
+Best,
+Derek Lankeaux
+[LinkedIn](https://linkedin.com/in/derek-lankeaux) ·
+[GitHub](https://github.com/dl1413) ·
+[Portfolio](https://dl1413.github.io/LLM-Portfolio/)
diff --git a/job_applications/2026-09-07/04_rag_applied_llm_engineer_remote.md b/job_applications/2026-09-07/04_rag_applied_llm_engineer_remote.md
new file mode 100644
index 0000000..1bbe0a8
--- /dev/null
+++ b/job_applications/2026-09-07/04_rag_applied_llm_engineer_remote.md
@@ -0,0 +1,42 @@
+# Cover letter — RAG / applied-LLM systems engineer (remote)
+
+**Archetype:** Remote-first applied-AI or product-AI teams building
+RAG assistants, agentic workflows, or knowledge platforms — from
+Series A/B startups (Elicit, Hebbia, Glean, Mendable, Vectara,
+LlamaIndex, Cohere applied side) through mid-stage vertical AI
+companies.
+
+---
+
+Dear {HIRING_MANAGER},
+
+I'm writing about the {ROLE_TITLE} role at {COMPANY}. My **RAG
+Production Pipeline** case study is the closest single project to
+what your team ships, and it's the one I'd point you to first.
+
+I designed a hybrid retrieval architecture — dense retrieval plus
+BM25, with a re-ranking stage, grounding checks, and confidence
+calibration on the generation side — and reported 96.3% Recall@10
+alongside 94.2% citation precision on the documented evaluation. The
+report goes past the headline metrics: latency and throughput budgets,
+drift monitoring, failure-mode analysis, and the observability,
+privacy, and rollback considerations I'd want in place before calling
+anything production-ready. That mindset carries through my
+**AI Safety Red-Team Evaluation** work — 12,500 response pairs, α =
+0.81, 96.8% held-out accuracy, with SHAP-backed risk analysis — and
+the **LLM Ensemble Textbook Bias Detection** project's 67,500-rating
+LLM-as-judge design.
+
+{ROLE_SPECIFIC_HOOK}
+
+On the stack: comfortable with Python, Qdrant, BM25, ColBERT, the
+OpenAI and Anthropic APIs, FastAPI, MLflow, Docker, and Kubernetes;
+comfortable enough with Kafka and Prometheus to reason about the
+operational surface. Fully remote works well for me, US-authorized,
+happy to walk through the RAG report in a first call.
+
+Best,
+Derek Lankeaux
+[LinkedIn](https://linkedin.com/in/derek-lankeaux) ·
+[GitHub](https://github.com/dl1413) ·
+[Portfolio](https://dl1413.github.io/LLM-Portfolio/)
diff --git a/job_applications/2026-09-07/05_enterprise_applied_data_scientist.md b/job_applications/2026-09-07/05_enterprise_applied_data_scientist.md
new file mode 100644
index 0000000..5733bf6
--- /dev/null
+++ b/job_applications/2026-09-07/05_enterprise_applied_data_scientist.md
@@ -0,0 +1,42 @@
+# Cover letter — Big-tech / enterprise applied data scientist
+
+**Archetype:** Meta, Google, Amazon, Microsoft, Netflix, Spotify, IBM
+Research, LinkedIn, Bloomberg, Adobe, Salesforce, Palantir applied-DS
+or research-engineer new-grad roles based in NYC or remote-eligible.
+
+---
+
+Dear {HIRING_MANAGER},
+
+I'm applying for the {ROLE_TITLE} role at {COMPANY}. I'm completing
+an M.S. in Applied Statistics at RIT (2026), and my portfolio is built
+to show that I can turn an ambiguous question into a defined study,
+run it with statistical rigor, and write it up so a partner team can
+act on the results.
+
+Two projects are the shortest path to seeing how I work. The
+**LLM Ensemble Textbook Bias Detection** case study evaluates 4,500
+passages with a rubric-based LLM-as-judge ensemble (67,500 ratings,
+Krippendorff's α = 0.84) and models publisher-level effects with
+Bayesian partial pooling in PyMC — the kind of hierarchical structure
+you need whenever you're comparing many small groups. The **AI Safety
+Red-Team Evaluation** project applies the same discipline to a harm-
+classification problem across 12,500 response pairs and six categories,
+reported with α = 0.81 and 96.8% held-out accuracy, and it separates
+annotation reliability, classifier performance, and Bayesian risk
+analysis as three distinct claims. The **RAG Production Pipeline**
+project (96.3% Recall@10; 94.2% citation precision) rounds out the
+systems side.
+
+{ROLE_SPECIFIC_HOOK}
+
+Stack: Python, SQL, R, scikit-learn, XGBoost, LightGBM, PyMC, MLflow,
+SHAP, plus the LLM tooling above. I'm comfortable owning a study from
+experimental design through the write-up, and I try to state limits
+in the same breath as results. NYC or remote both work; US-authorized.
+
+Best,
+Derek Lankeaux
+[LinkedIn](https://linkedin.com/in/derek-lankeaux) ·
+[GitHub](https://github.com/dl1413) ·
+[Portfolio](https://dl1413.github.io/LLM-Portfolio/)
diff --git a/job_applications/2026-09-07/README.md b/job_applications/2026-09-07/README.md
new file mode 100644
index 0000000..cf1163a
--- /dev/null
+++ b/job_applications/2026-09-07/README.md
@@ -0,0 +1,17 @@
+# Applications — 2026-09-07
+
+Five tailored cover-letter drafts for today's applications, all pitched
+on the same three primary projects listed in
+[`../README.md`](../README.md).
+
+| # | Archetype | File |
+|---|---|---|
+| 1 | Frontier AI lab — safety / model evaluations | [`01_frontier_ai_lab_safety_evaluations.md`](./01_frontier_ai_lab_safety_evaluations.md) |
+| 2 | NYC fintech / quant — data scientist | [`02_nyc_fintech_data_scientist.md`](./02_nyc_fintech_data_scientist.md) |
+| 3 | LLM evaluation / observability startup | [`03_llm_evaluation_startup.md`](./03_llm_evaluation_startup.md) |
+| 4 | RAG / applied-LLM systems engineer (remote) | [`04_rag_applied_llm_engineer_remote.md`](./04_rag_applied_llm_engineer_remote.md) |
+| 5 | Big-tech / enterprise applied data scientist | [`05_enterprise_applied_data_scientist.md`](./05_enterprise_applied_data_scientist.md) |
+
+Fill placeholders (`{COMPANY}`, `{ROLE_TITLE}`, `{HIRING_MANAGER}`,
+`{ROLE_SPECIFIC_HOOK}`) per opening, then log to
+[`tracker.md`](./tracker.md).
diff --git a/job_applications/2026-09-07/tracker.md b/job_applications/2026-09-07/tracker.md
new file mode 100644
index 0000000..915f8f8
--- /dev/null
+++ b/job_applications/2026-09-07/tracker.md
@@ -0,0 +1,11 @@
+# Submission tracker — 2026-09-07
+
+Target: 5 applications, NYC or remote.
+
+| # | Archetype | Company | Role | Posting URL | Sent (Y/N) | Notes |
+|---|---|---|---|---|---|---|
+| 1 | Frontier AI lab — safety / evals |  |  |  |  |  |
+| 2 | NYC fintech / quant — data scientist |  |  |  |  |  |
+| 3 | LLM evaluation / observability startup |  |  |  |  |  |
+| 4 | RAG / applied-LLM systems engineer |  |  |  |  |  |
+| 5 | Big-tech / enterprise applied data scientist |  |  |  |  |  |
diff --git a/job_applications/README.md b/job_applications/README.md
new file mode 100644
index 0000000..0264d9e
--- /dev/null
+++ b/job_applications/README.md
@@ -0,0 +1,42 @@
+# Job applications
+
+Daily packages of tailored application materials for Derek Lankeaux's 2026
+data science / applied ML / LLM evaluation job search. Each dated folder
+contains five cover-letter drafts targeting distinct role archetypes in
+New York City or remote roles, all built around Derek's three primary
+portfolio projects.
+
+## Three primary projects
+
+The résumé's "Data Scientist | Applied Statistician | LLM Evaluation &
+Applied ML" positioning is anchored by three projects that carry across
+every letter in this folder:
+
+1. **AI Safety Red-Team Evaluation** — LLM ensemble labels + supervised
+   harm classification across 12,500 response pairs; α = 0.81; 96.8%
+   held-out accuracy; Bayesian risk analysis with PyMC and SHAP.
+2. **LLM Ensemble Textbook Bias Detection** — rubric-based LLM-as-judge
+   evaluation of 4,500 passages yielding 67,500 ratings; α = 0.84;
+   Bayesian partial pooling for publisher-level effects.
+3. **RAG Production Pipeline** — hybrid dense + BM25 retrieval with
+   re-ranking, grounding checks, and confidence calibration; 96.3%
+   Recall@10; 94.2% citation precision; latency, drift, and failure-mode
+   analysis.
+
+The Breast Cancer Classification benchmark stays in the portfolio for
+roles that specifically value diagnostic-ML rigor; it isn't featured
+in the daily archetype letters because those target LLM/AI-forward
+employers where the three primary projects tell a tighter story.
+
+## How to use each daily folder
+
+1. Pick five real openings for the day (NYC or remote, junior/new-grad
+   data science, applied ML, or LLM evaluation).
+2. For each opening, choose the letter whose archetype fits best and
+   fill in the `{COMPANY}`, `{ROLE_TITLE}`, `{HIRING_MANAGER}`, and any
+   `{ROLE_SPECIFIC_HOOK}` placeholders — one or two sentences of
+   company-specific evidence you pulled from the job post.
+3. Log the submission in the day's `tracker.md`.
+
+Letters are drafts, not autofilled applications. Derek reviews and sends
+each one himself.