na-tech-jobs

A working ML platform for the senior data-science / ML hiring market in the US and Canada. Weekly ingest from public ATS APIs, four trained models, hybrid retrieval, and an LLM analytics layer. Runs on a $9/month Hugging Face Space.

I started building this for my own job search. Each piece exists because I had a question I couldn't answer with LinkedIn or a spreadsheet: what's the actual salary distribution for senior MLE roles in Toronto, which companies are hiring at staff level right now, where do my skills line up with what's open. The ingestion cron, the salary regressor, the seniority and role-family classifiers, the matcher, the analytics tab — each one is the answer to one of those questions.

🔗 Live links

🚀 Demo Space (dark theme, 5 tabs)	https://arjun10g-na-tech-jobs.hf.space
📦 Source code	https://github.com/Arjun10g/na-tech-jobs
📊 Dataset (12,334 active jobs, weekly)	https://huggingface.co/datasets/arjun10g/na-tech-jobs
🧠 Models on the Hub	salary · seniority · role_family · skills

What's live on the Space

Five tabs, all hitting real data and real model output.

Salary. Paste a job description, the regex cascade extracts ~20 features, an XGBoost regressor predicts the maximum salary in USD/year. Edit any extracted field and the prediction updates. Held-out test-MAE is $29,091, MAPE 14.7%.

Search. Substring match on title and company, with country and role-family dropdowns. It's the boring tab — useful for spot-checking the corpus before reaching for the matcher.

Matcher. Natural-language query (or a pasted resume blurb) goes through dense Qdrant retrieval, an optional cross-encoder rerank, and parent-chunk hydration. You get a ranked table of jobs with a clickable apply link, plus a short LLM-generated paragraph explaining which 1-2 jobs best fit and why. Recall@10 against a labeled query set is 0.486 with rerank, 0.363 without.

Analytics. Plain-English question → Qwen2.5-7B writes DuckDB SQL → sqlglot-based safety layer rejects anything that isn't a read-only SELECT over the allowlist → DuckDB executes against the enriched parquet. The executed SQL is shown next to the result so you can verify what actually ran. 61 tests pin the safety contract; 4 out of 4 sample questions returned the right answer in a live smoke test.

Dashboard. Pipeline health (last ingest, last enrichment, per-extractor counts), market-trend tables (salary by role × seniority, top employers, role-family share by country, top skills), and the latest drift report. Refresh button pulls the curated parquet from the Hub on first call.

---

Architecture

flowchart TB
    subgraph Ingest["⏱ Weekly — Sundays 02:00 UTC"]
        A1[Greenhouse]
        A2[Lever]
        A3[Ashby]
        A4[Workable]
        A5[SmartRecruiters]
        A6[Workday tenants]
        A1 & A2 & A3 & A4 & A5 & A6 --> ORCH[orchestrator<br/>asyncio + httpx]
        ORCH --> NORM[normalize.py<br/>title/loc/currency/period]
        NORM --> DEDUP[dedup vs prior snapshot]
        DEDUP --> QUAL[Pandera schema validation]
        QUAL --> SNAP[(snapshots/&lt;date&gt;/jobs.parquet)]
    end

    SNAP --> CURATE[curated/build.py<br/>DuckDB]
    CURATE --> CURATED[(curated/jobs.parquet<br/>12,334 active jobs)]

    subgraph Features["📦 Phase 1b — feature extraction cascade"]
        CURATED --> RGX[Tier 1 regex extractors<br/>seniority, role, skills,<br/>salary, sponsorship, ...]
        RGX -. opt-in .-> NUE[Tier 2 NuExtract LLM<br/>monthly retrain only]
    end

    subgraph Models["🧠 Phase 2-4 — models"]
        CURATED --> XGB[XGBoost salary regressor<br/>test-MAE $29,091]
        CURATED --> SENMODEL[MiniLM + LR seniority<br/>val f1_macro 0.812]
        CURATED --> ROLEMODEL[MiniLM + LR role_family<br/>val f1_macro 0.934]
    end

    XGB & SENMODEL & ROLEMODEL --> ENRICH[curated/enrich.py<br/>versioned predictions]
    ENRICH --> ENRICHED[(curated_enriched/jobs.parquet<br/>seniority_label_v1, role_family_v1,<br/>predicted_salary_usd_v1, ...)]

    subgraph RAG["🔍 Phase 5-6 — retrieval"]
        ENRICHED --> CHUNK[parent-child chunking<br/>29k parents, 120k children]
        CHUNK --> EMB[MiniLM dense<br/>bge-m3 v1.1 reindex]
        EMB --> QDRANT[(Qdrant local-mode<br/>jobs_dense + jobs_multivec)]
        QDRANT --> RAGPIPE[hybrid pipeline<br/>RRF + cross-encoder rerank<br/>recall@10 = 0.486]
    end

    subgraph App["🛰 HF Space — always-on"]
        ENRICHED --> SALARYTAB[Salary tab]
        ENRICHED --> SEARCHTAB[Search tab]
        RAGPIPE --> MATCHERTAB[Matcher tab]
        ENRICHED --> ANALYTICSTAB[Analytics tab<br/>NL→SQL + sqlglot safety]
        ENRICHED --> DASHTAB[Dashboard tab<br/>drift + market trends]
    end

    subgraph Ops["🔁 Closed-loop ops"]
        SNAP & ENRICHED --> DRIFT[Mon 03:00 UTC drift cron<br/>Evidently PSI]
        DRIFT -. PSI ≥ 0.20 .-> RETRAIN
        DRIFT --> ALERT[Discord webhook]
        RETRAIN[1st @ 04:00 UTC monthly<br/>champion/challenger gate] --> XGB & SENMODEL & ROLEMODEL
        RETRAIN --> ALERT
    end

    classDef artifact fill:#fef3c7,stroke:#d97706,color:#000
    classDef live fill:#dbeafe,stroke:#2563eb,color:#000
    classDef ops fill:#fee2e2,stroke:#dc2626,color:#000
    class SNAP,CURATED,ENRICHED,QDRANT artifact
    class SALARYTAB,SEARCHTAB,MATCHERTAB,ANALYTICSTAB,DASHTAB live
    class DRIFT,RETRAIN,ALERT ops

It's one loop. Ingest pulls from ATS APIs every Sunday, the curated layer dedups and validates, the enrichment script scores every job with the four models and writes a parquet with versioned columns (predicted_salary_usd_v1, seniority_label_v1, …). The chunker + embedder + Qdrant index serve the matcher; the same enriched parquet backs analytics and the dashboard. Drift detection runs every Monday; the retrain cron runs on the 1st of each month with a champion/challenger gate. Everything's versioned through git and HF Hub commits. Detailed architecture in CLAUDE.md §4 + §8.

---

Headline numbers

Surface	Metric	Value
Dataset	Active jobs in latest snapshot	12,334
	Companies	477
	Salary disclosure rate	49.8%
	Median disclosed / predicted salary (USD/yr)	$195k / $187.5k
Salary regressor	Test-MAE (XGBoost + Optuna 50)	$29,091 (95% CI $27k–$31k)
	Test-MAPE	14.7%
	R² log-salary	0.730
Seniority classifier	f1_macro vs reviewed gold	0.812 (95% CI [0.73, 0.87])
Role-family classifier	f1_macro vs reviewed gold	0.934 (95% CI [0.88, 0.98])
Hybrid retrieval	recall@10 (hybrid+rerank, 48 labeled queries)	0.486 (vs dense 0.363)
	MRR	0.518
NL→SQL	Live smoke set accuracy	4/4
	Safety-layer test count	61
CI	Total tests passing	371

---

Phase status

Phase	Status	Headline
0–1 — Scaffold + Ingestion	✅	Repo + CI + 65 ATS handles → weekly parquet → HF Dataset (~12.3k jobs)
2 — Salary regressor	✅	Six-tier ladder; Tier 5 XGBoost test-MAE $29,091, every CI cleanly excludes the previous
3 — First deployable	✅	Salary prediction + curated search live on the Space
4 — Multi-model + payload enrichment	✅	MiniLM + LR classifiers (seniority val 0.812, role_family 0.934), regex skills, all 12,334 jobs enriched with versioned predictions
5 — Retrieval stack	✅	Parent-child chunking (29k/120k) + Qdrant + dense (MiniLM) hybrid + cross-encoder rerank + Matcher tab; bge-m3 reindex queued v1.1
6a — Retrieval eval	✅	48-query labeled set; hybrid+rerank recall@10 = 0.486 (+34% vs dense)
7 — NL→SQL analytics	✅	Mandatory sqlglot safety layer + 61 tests; Anthropic / HF Inference / mock LLM backends; Analytics tab live
8a — Dashboard	✅	Drift detection (PSI) + pipeline-health + market-trend tabs
8b — CI workflows	✅	drift.yml (Mondays 03:00 UTC) + retrain.yml (monthly, champion/challenger gate)
9 — Polish	✅	Mermaid architecture diagram, dark UI, HF dataset YAML frontmatter, model-card cross-links

---

Salary regressor — six-tier ladder

All evaluated on the same frozen 80/20 train/test split (n_test=1,226), with bootstrap 95% CIs on test-MAE and 5-fold CV-MAE on the training set as a generalization sanity check.

Tier	Test-MAE	Test 95% CI	CV-MAE	CV 95% CI	MAPE	R² log
0 constant baseline	$60,509	$57k–$64k	$62,279	$61k–$64k	33.7%	≈0
1 stratified mean	$59,589	$56k–$63k	$61,623	$60k–$63k	32.9%	0.045
2 Mincer OLS	$51,322	$48k–$54k	$52,041	$50k–$53k	27.4%	0.283
3 Ridge (full encoder)	$43,199	$41k–$45k	$42,179	$41k–$43k	23.3%	0.462
4 Random Forest	$35,935	$34k–$38k	$37,016	$36k–$38k	19.0%	0.615
5 XGBoost + Optuna(50)	$29,091	$27k–$31k	$30,533	$29k–$32k	14.7%	0.730

Each tier's 95% bootstrap CI sits cleanly above the previous tier's; test-MAE and CV-MAE agree to within 5%, so this isn't a lucky test draw or an overfit. The CLAUDE.md target of MAE under $25k isn't met yet — closing that gap is the bge-m3 description-embedding work in v1.1.

Methodology in LITERATURE_REVIEW.md §16: parsimonious-first ladder informed by Breiman's two-cultures essay, the Mincer earnings function, and the recent gradient-boosting-on-tabular results from Shwartz-Ziv & Armon (2022) and Grinsztajn et al (2022).

---

Title classifiers — frozen MiniLM + multinomial LR

Classifier	Classes	Train rows	f1_macro vs regex	f1_macro vs reviewed gold	95% CI	Repo
seniority	7 (intern…director)	6,361	0.831	0.812	[0.73, 0.87]	...-seniority-v1
role_family	6 (DS / DA / DE / MLE / RS / AS / SWE-ML)	569	0.915	0.934	[0.88, 0.98]	...-role_family-v1

Architecture is frozen sentence-transformers/all-MiniLM-L6-v2 (22M params, 384-dim) for embeddings, sklearn multinomial LR for the head. Class weights balanced. C picked by 5-fold stratified CV from {0.1, 1, 10}.

Why a linear probe and not CLAUDE.md §7's locked DeBERTa-v3 + LoRA? For short-text classification on weakly-supervised labels, a linear probe over a strong general-purpose embedder lands at the same operating point for roughly two orders of magnitude less compute. The literature that motivated the call is in LITERATURE_REVIEW.md §17: Peters et al (2019) on when fine-tuning helps, Tunstall et al's SetFit, Joulin's FastText. v1.2 will run the DeBERTa-v3 comparison against the human-reviewed gold once that test set lands.

Training labels come from the regex extractors. Rows where the regex fell back to its default (mid / Other / Manager) are dropped from training because the fallback is too noisy. The "vs regex" column measures held-out agreement against the same regex labels; "vs reviewed gold" comes from a two-pass Claude-reviewed sample (230 rows per classifier, first-pass labelers then second-pass reviewers shown the proposal + the classifier's prediction). The reviewer override rate was 1.7%. Process notes are in MAINTENANCE.md.

Skills are regex-first by default. extracted_skills_v1 gets populated from ingestion/feature_extraction/regex/tech_stack.py on every weekly ingest — about 70 canonical names, 64.7% coverage, runs in milliseconds, and (most importantly) is free. NuExtract (arjun10g/na-tech-jobs-skills-v1) is wired but opt-in via --skills-mode=nuextract. It runs during the monthly retrain so the enriched skills column gets the LLM-tier output once a month.

curated_enriched/jobs.parquet on the Hub has all 12,334 jobs scored with the versioned columns: seniority_label_v1, seniority_confidence_v1, role_family_v1, role_family_confidence_v1, predicted_salary_usd_v1, extracted_skills_v1, prediction_model_version.

uv run python -m curated.enrich --push-to-hub  # rebuild + push

---

Hybrid retrieval (Phase 5)

	Detail
Chunking	parent-child RecursiveCharacterTextSplitter (~1024-tok parents, ~256-tok / 32-overlap children, hierarchical markdown separators)
Volume indexed	12,334 jobs → 29,311 parents → 120,004 children
Embedder (v1)	sentence-transformers/all-MiniLM-L6-v2 (384-dim, dense-only). bge-m3 (1024-dim dense + sparse + ColBERT multivec) wired but reindex deferred to v1.1
Vector store	Qdrant local-mode at data/qdrant/. Two collections: jobs_dense (named dense + sparse vectors, HNSW + int8 scalar quantization) and jobs_multivec (ColBERT MaxSim, populated by v1.1)
Pipeline	dense first-pass (top 100) → optional sparse search → RRF fusion (k=60) → optional cross-encoder rerank (bge-reranker-v2-m3 or lite ms-marco MiniLM) → parent-chunk hydration → top-K → LLM rationale
Index time	14:05 wall-clock for the full MiniLM index on Apple MPS (142 chunks/sec)
Latency	<1 s end-to-end without rerank; ~3-5 s with rerank enabled

A note on how the index gets to the Space: data/ is gitignored and excluded from the deploy. The Qdrant directory ships out-of-band as a gzipped tarball at arjun10g/na-tech-jobs/qdrant/qdrant_minilm_v1.tar.gz (~260 MB). On the Space's first matcher request, app/retriever_loader.py downloads and extracts it. Subsequent requests hit the local directory. The same pattern will work for the bge-m3 tarball when it lands.

Retrieval eval

48 labeled queries: 30 are sampled job titles with the gold pool defined by all jobs sharing the same normalized title and country. The other 18 are hand-written role+seniority queries with a gold pool defined by the classifier-label slice on the enriched parquet.

Variant	recall@5	recall@10	recall@20	MRR	nDCG@10	latency
dense	0.291	0.363	0.393	0.412	0.349	186 ms/q
hybrid+rerank	0.421	0.486	0.511	0.518	0.476	700 ms/q

The cross-encoder is the biggest single quality lever in v1: +34% on recall@10 for ~4x the latency. Hybrid here is identical to dense because the MiniLM index has no sparse vectors yet; the bge-m3 reindex adds the sparse leg, and the same indexer already wires the ColBERT multi-vec collection. HyDE (a Qwen-generated hypothetical doc fed to retrieval before the real query) is queued behind the bge-m3 reindex since both rows benefit from sparse search.

# dev / fast iteration (MiniLM, ~14 min for 120k chunks)
uv run python -m scripts.index_jobs --lite --force-recreate
# production (bge-m3 dense + sparse, ~6-8 hr on MPS or ~30 min on A10G HF Job)
uv run python -m scripts.index_jobs --force-recreate

# eval
uv run python -m eval.run_retrieval_eval --variants dense hybrid+rerank

---

NL→SQL analytics

The Analytics tab takes a plain-English question and runs DuckDB SQL over the enriched parquet. The interesting work isn't in the LLM call, it's in the safety layer that sits between the LLM and the database. Six things have to pass before the SQL is allowed to run:

Layer	What it does
Pre-parse keyword filter	Word-boundary scan rejects INSERT/UPDATE/DELETE/DROP/CREATE/ALTER/COPY/ATTACH/PRAGMA/INSTALL/LOAD/... before the parser runs. Conservatively catches keyword-in-string-literal attacks.
sqlglot single-statement parse	One statement, must be a Select or Subquery.
Table allowlist	Only jobs is queryable; CTE-introduced names admitted. Disallowed tables in JOIN clauses caught.
Column allowlist	Per-table allowlist of ~40 columns. SELECT-list aliases and CTE columns admitted; everything else rejected.
Row + time caps	DuckDB statement_timeout=5s + outer LIMIT 1000.
Always-show SQL	Executed SQL rendered alongside results so the user verifies what actually ran.

default_llm() picks a backend automatically: Anthropic Claude when ANTHROPIC_API_KEY is set, HF Inference / Qwen2.5-7B-Instruct when only HF_TOKEN is, fail-loud otherwise. Tests use a MockLLM so CI never depends on a live API.

61 safety-layer tests lock the contract. Every denylisted keyword is parametrized, multi-statement payloads get caught, every legal SELECT shape (joins, CTEs, subqueries, aggregates with aliases used in ORDER BY / HAVING) passes, every illegal shape gets rejected.

Live smoke test against Qwen2.5-7B over the enriched parquet:

Question	Result
How many senior MLE jobs are open in the US right now?	276
Top 5 companies hiring data scientists, ranked by number of postings.	Pinterest 65, Robinhood 61, Databricks 47, Whatnot 45, Jane Street 23
Average disclosed salary range for staff-level roles?	$212,863 USD/yr
Distribution of role_family_v1 across countries.	12 rows; US dominates DE (3,492), DA (2,918), SWE-ML (2,473); CA SWE-ML-heavy at 29.6%

uv run python -m rag.nl2sql "median predicted salary by country for senior MLEs"

---

Operations

Four crons close the loop. Discord webhooks fire on success and failure.

Cron	When	What
ingest.yml	Sun 02:00 UTC	Pull every ATS, dedup vs prior, validate, push snapshot to dataset repo, Discord alert on failure
drift.yml	Mon 03:00 UTC	Compare latest vs 4-week-old snapshot via Evidently. PSI ≥ 0.20 on any tracked feature → priority breach → flag retrain + Discord alert
retrain.yml	1st @ 04:00 UTC	Matrix over [seniority, role_family]. Pull champion training_summary.json, train challenger, apply monitoring.champion_challenger gate (primary +1%, no secondary > -2%), publish only if promoted
deploy-space.yml	On push to main	Generate requirements.txt from space-runtime extras, rsync runtime files, push to HF Space, write Gradio frontmatter README

The dashboard tab pulls whichever drift report is newest in reports/drift/<date>.html and renders it inline next to a metrics card. The pipeline-health card reads ingestion_stats.json from the latest snapshot directory.

---

Project documents

File	What it contains
CLAUDE.md	Project bible — architecture, locked decisions, phased plan, risks
DATA_DICTIONARY.md	Every column in data/curated/jobs.parquet — type, fill rate, predictor decision
LITERATURE_REVIEW.md	~1k-line predictor-by-predictor review with 50+ citations + ideal-EDA self-audit
MAINTENANCE.md	Running known-issues / debt log, resolved entries kept for project history
eda/reports/<date>/report.md	Statistical audit + 11 plots per snapshot

---

Quickstart

# clone + install Python 3.11 + dev deps
uv sync --group dev

# run the local Gradio app (loads model + curated parquet from HF Hub)
uv run python -m app.main

# run lint + tests
uv run ruff format --check .
uv run ruff check .
uv run pytest

# smoke ingest (5 companies, no HF push)
uv run python -m ingestion.orchestrator --output-dir data --limit 5

# full ingest with HF Dataset push (needs HF_TOKEN)
uv run python -m ingestion.orchestrator --output-dir data --push-to-hub --alert

# rebuild curated layer + enriched predictions
uv run python -m curated.build --push-to-hub
uv run python -m curated.enrich --push-to-hub

# retrain a classifier + push to HF Hub
uv sync --extra ml --group dev
uv run python -m models.seniority.train  # ~30 sec on CPU
uv run python -m scripts.publish_classifier seniority --create

# rebuild Qdrant index + publish tarball
uv run python -m scripts.index_jobs --lite --force-recreate
tar -cf - -C data qdrant | gzip -9 > qdrant_minilm_v1.tar.gz
uv run python -m scripts.publish_qdrant_index --tarball qdrant_minilm_v1.tar.gz

Optional dependency groups

Group	Purpose	Install
ml	Training stack (torch, transformers, xgboost, mlflow, sklearn, optuna)	uv sync --extra ml
eda	matplotlib, seaborn, statsmodels, missingno, tabulate	uv sync --extra eda
space-runtime	Lean runtime deps for the Space (xgboost, sklearn, sqlglot, qdrant-client, sentence-transformers, anthropic)	uv sync --extra space-runtime
rag	Vector store, chunking, NL→SQL, PDF parsing	uv sync --extra rag
monitoring	Evidently for drift reports	uv sync --extra monitoring
api	FastAPI for programmatic endpoints	uv sync --extra api

Repository layout

app/                  Gradio front end (5 tabs, dark-themed)
ingestion/            ATS extractors + feature cascade
curated/              DuckDB layer over weekly snapshots, model-prediction enrichment
eda/                  Statistical audit + train/test split
models/               salary, seniority, role_family, skills, embeddings
rag/                  Chunking, embedder, qdrant client, reranker, pipeline, nl2sql
monitoring/           Drift, pipeline health, market trends, champion/challenger gate
eval/                 Retrieval-eval harness, labeled queries, classifier test sets
scripts/              Publish models/dataset, build query sets, index_jobs, etc.
.github/workflows/    CI, weekly ingest, drift cron, monthly retrain, deploy-space
tests/                371 tests across all surfaces

Full structure in CLAUDE.md §9.

Secrets

Copy .env.example to .env and fill in HF_TOKEN (with both Hub-write and Inference-Provider permissions if you want the Analytics tab to use Qwen) and DISCORD_WEBHOOK_URL for ingest/drift/retrain alerts. ANTHROPIC_API_KEY is optional and preferred over HF Inference when set.

License

MIT. Models and datasets licensed individually — see their respective cards.