AI PM Job Radar
Every product, program and technical program manager opening at 36 AI and tech companies, sorted by a decision model in 15 seconds for about 11 cents.
I built this for my own search: I'm a Toronto-based program manager targeting AI program, TPM and AI product roles. Job boards can't answer the questions I actually care about: is the product itself AI, does the role require hands-on ML, is it a program role or a product role, and can I do it from Toronto. A title filter gets these wrong, and running an LLM over a thousand postings every day is slow and expensive. So the radar splits the work: code handles the facts, and Jev (TypeSafe's System One model) handles the meaning.
Live: ai-pm-job-radar.vercel.app, every labeled role with filters for bucket, company and Toronto fit.
Latest run (2026-10-24)
The live page refreshes every week in GitHub Actions (see .github/workflows/weekly-refresh.yml). These are the numbers from the latest run, next to the first scheduled run (the ledger on the page starts there too).
| 2026-10-24 | First scheduled run, 2026-10-05 |
|---|
| Open roles scanned | 8,387 across 36 public job boards | 8,427 |
| Product or program titles | 981 | 1,016 |
| Labeled by Jev | 981 in 6.6 s (p50 95 ms per call), 0 errors | 1,016 in 13.5 s (p50 176 ms), 0 errors |
| Cost | $0.111 (2.64M input tokens at $0.042 per million; output is free) | $0.107 (2.55M input tokens) |
| Model | jev-1.14.0 | jev-1.13.0 |
| Bucket | 2026-10-24 | 2026-10-05 |
|---|
| AI program / TPM | 105 | 127 |
| AI PM, no ML background required | 160 | 184 |
| AI PM, ML background required | 16 | 11 |
| Program / TPM, not AI | 134 | 145 |
| PM, not AI | 132 | 113 |
| Not a PM or program role | 376 | 400 |
| Low confidence, needs review | 58 | 36 |
Of the 981 roles, 53 can be done from Toronto (remote in Canada, or a Toronto, Ontario or other Canadian office), and 21 of those are PM or program roles (first scheduled run: 27 of 58).
Finding: only 16 of 176 AI PM roles list hands-on ML experience as a must-have (first scheduled run: 11 of 195). Most AI product roles want judgment about AI, not a model-building background.
How it works
radar/boards.json which companies (Greenhouse, Ashby, Lever slugs)
radar/fetch.py pulls every open role, keeps product/program titles, computes facts in code
radar/spec.json the questions Jev answers per role, plus bucket rules
scripts/jev.py runs the spec over every role, caches answers, writes CSV + summary
scripts/export_site.py writes site/data.json (posting facts + labels only, no descriptions)
site/ static page on Vercel: no server, no API key, strict CSP
eval/ 40 hand-labeled postings, the scorer, and RESULTS.md
Code computes facts. Pay ranges, the remote flag and where a Toronto-based person stands (remote_canada, toronto_ontario_office, canada_other_city, canada_unspecified, outside_canada) come from regex over the posting. The first version asked Jev for location, and it labeled four "Vancouver, BC" roles as US-only. A location string is a fact, so I moved it into code.
Jev judges meaning. There are six typed questions per posting, each answered with a probability:
| Question | Type | Why it can't be a keyword rule |
|---|
role_kind | choice of 9 | "Product Marketing Lead" and "Fellows Program" contain "product" and "program" but aren't PM or program jobs |
ai_scope | choice of 4 | Every posting at an AI company mentions AI; the question is whether this team's product is AI |
seniority | 5-level score | Titles are inconsistent across companies |
requires_ml_background | yes/no | Separates "must have shipped ML models" from "comfortable discussing AI" |
requires_hands_on_ai | yes/no | Catches roles that screen for building with AI tools or agents, or required AI fluency, without asking for an ML background |
requires_engineering_background | yes/no | Separates a CS-degree requirement from working with engineers |
Buckets combine the answers in code with a confidence floor of 0.6, and anything below it goes to review. All questions and thresholds live in radar/spec.json.
Design decisions
- Escape options on every choice (
other, unclear), so the model is never forced to guess.
- Company boilerplate is excluded by instruction.
ai_scope tells Jev to ignore mission statements and benefits text that appear on every posting.
- Iteration was measured on a sample first. Each change was tested on a 40-role random sample (about half a cent) and a 55-role Canada set before the full run. One sample error, an "AI & LLM conversational fluency" line scored as a hard ML requirement, led to a sharper
requires_ml_background definition.
Evaluation
I labeled 40 postings by hand and scored Jev's answers against mine (eval/, full results in eval/RESULTS.md). The sample takes a fixed number of postings from each bucket, so rare buckets get enough rows to check. These are per-bucket numbers, not the accuracy on a random posting.
| All 40 | Labeled blind (21-40) |
|---|
| In an AI target bucket or not (rows not sent to review) | 32/36 | 15/18 |
| Exact bucket (rows not sent to review) | 26/36 | 11/18 |
AI vs not AI (ai_scope, native and platform merged) | 33/40 | 16/20 |
role_kind | 31/40 | 15/20 |
| Seniority within one level | 40/40 | 20/20 |
requires_ml_background (strict) | 36/40 | 17/20 |
requires_hands_on_ai | 31/40 | 14/20 |
requires_engineering_background | 34/40 | 17/20 |
Model jev-1.13.0. Answers come from the 2026-10-04 run, except requires_hands_on_ai, which comes from a 2026-10-05 re-run on the same 40 postings ($0.005).
How the labels were made. Postings 21-40 were labeled without seeing Jev's answers. Postings 1-20 were labeled the same way, then 9 of their labels were revised after I saw Jev's answers and settled what "AI scope" means: this team builds AI, not this is an AI company. The originals are kept in eval/labels_batch1_before_review.csv. Treat the blind column as the honest one.
What the eval changed. On requires_ml_background, 12 of the first disagreements were me saying yes where Jev said no. The posting text showed I was answering a broader question: does the role need hands-on AI skills? That's a different question from whether it needs an ML background, so the spec now asks both. requires_hands_on_ai was added after this, and its wording was written after I had seen my own labels, so treat its 31/40 as a first read rather than a held-out score. I then relabeled the strict ML question for the 16 postings where I'd said yes.
Where Jev and I still disagree. Most misses don't change the bucket. They're TPM vs program manager, AI-native vs AI-platform, and seniority off by one level. The ones that do change the bucket are borderline AI-scope calls and 4 strict-ML calls that move a posting between the two AI PM buckets, plus the 4 postings Jev sent to review, which the table leaves out.
Repeatability. Re-running Jev on the same 40 postings gave the same role_kind and ai_scope on all 40, and the same bucket on 38. Both changes were near a cutoff: an ML score of 0.53 became 0.49, and a confidence of 0.63 became 0.54.
Run it yourself
Python 3.9+, standard library only.
cp .env.example .env.local # paste a TypeSafe key: https://console.typesafe.ai/keys
python3 scripts/jev.py check
mkdir -p jev-runs/radar && cd jev-runs/radar
python3 ../../radar/fetch.py --out input.jsonl
python3 ../../scripts/jev.py run ../../radar/spec.json input.jsonl --out . --sample 40 # under a cent
python3 ../../scripts/jev.py run ../../radar/spec.json input.jsonl --out . # everything
The site refreshes every week in GitHub Actions (.github/workflows/weekly-refresh.yml): fetch, label (capped at $0.30), export, push, and Vercel redeploys. The TypeSafe key is an encrypted repository secret; a run can also be started from the Actions tab. scripts/weekly_refresh.sh does the same refresh from a local machine with the key in .env.local. The export refuses to publish if more than 5% of labels failed or the role count halves. To refresh by hand, export and push:
python3 scripts/export_site.py jev-runs/radar/full --scanned <total printed by fetch.py>
git add site/data.json && git commit -m "Refresh radar data" && git push
--dry-run prints the token and cost estimate without calling the API. Reruns only pay for new postings. jev.py has an untested Vercel AI Gateway path that is used only when AI_GATEWAY_API_KEY is set and no TypeSafe key is.
Credits
Built on the how-to-jev skill by Aakash Gupta (product-growth.com, MIT): the jev.py runner, validator and the original job-radar recipe. My changes: program and TPM roles, Toronto/Canada location logic, the tightened ML question, Vercel AI Gateway support, and the bucket design for my search.
