About · Deep dive · a working model on sample data

Example · Sample data. The scores were computed ahead of time.

The candidate, Morgan Vale, is a made-up sample person: product leader in logistics, travel, energy and learning software.

What it is

A product I built that finds the roles that best match a record, and tracks each one.

JobWatch searches open roles from many sources, matches them against 17 years of experience held in RAG and a vector index, scores the fit with reasons, and tracks applications and their outcomes. The result: selective, best-fit applications.

Who it is for

A candidate who spends hours reading postings that do not fit.

The time goes into searching and reading, not into the few roles that fit best. The tool does the reading first and shows its reasons, so the candidate's time goes where the fit is.

The outcome

Find relevant roles faster, and see why each one fits.

Each role gets a fit score, with the reasons behind it.
It tracks applications and interviews, each with its status from review to outcome.
The interview line-up sits in one place: every role In interview, with its notes.
Feedback notes on each role record what was learned, to improve the next search.

How it works

JobWatch searches, matches and scores open roles and tracks applications, so the candidate applies selectively, only to the best fits.

JobWatch searches open roles from several sources and keeps one list, each role once.
JobWatch matches the roles against 17 years of experience held in RAG and a vector index, retrieving the nearest passages.
A separate AI assessment scores the fit, with reasons.
The candidate tracks applications in six statuses and records each outcome; reports show the funnel and each source's yield.
The result: the candidate applies selectively, only to the best fits.

Explanatory graphic. Five steps; with scripts off or reduced motion the finished diagram is shown with the steps listed.

When it was built

I designed it and built it myself, in 2026.

Work-history passages are embedded with Vertex AI gemini-embedding-001 and stored in a Firestore vector index; hybrid retrieval joins vector and exact-term results by reciprocal-rank fusion. SQLite and BigQuery hold the scan's working state and a reporting copy.

Why I built it

To read the postings so the candidate does not have to, and to show the reasons.

A score without reasons cannot be trusted; a list without a record cannot be learned from. The model shows both: the reasons behind each score and the reports on the tracked outcomes.

So what

Evidence that I build with AI myself, with the candidate choosing.

The candidate chooses the roles and applies.

Preserve the candidate decision

A score needs reasons.

Keep fit assessment, missing evidence, no supported fit, empty results and unavailable integrations distinct. The live state label below follows this sample model’s actual selection and filters; it does not fabricate a score or replace a role.

What changed since the older sample model?

The older public model describes vector similarity as part of matching. The repository uses vector retrieval to find evidence for documents rather than to compute the fit score. Role assessment and document-evidence retrieval therefore remain separate.

The existing 40 roles, 25 history entries, precomputed scores, reasons and retrieval passages are preserved. This older local sample has six status options. The repository has seven visible states: To review, Ready to apply, Ready to submit, Applied, In interview, Rejected and Not interested; Closed is hidden by default. These source capabilities do not add a new state to this sample or validate ranking quality. Laptop Gmail intake is documented separately from the proposed unified MCP flow; this preview does not execute tasks or grant integration permissions.

Static sample

The complete sample role list and explanations are readable below. Interactive state updates require JavaScript.

The model

Sample data. The scores were computed ahead of time. Everything here is made up; state lives in memory only and resets on reload.

Loading example data…

Integration availability in this example

Live role search, email/MCP and calendar connections are unavailable in this example. The role list and history come from the supplied sample data. Status and notes stay in this page's memory and reset on reload.

Strong and weak example fits

The strong example, VP Product at Quernlade Logistics, scores 96 with four requirements supported by experience passages. The weak example, Principal Product Manager at Orpinecote Field Services, scores 39: team leadership has passage 4, while enterprise sales and robotics have no supporting passage. Its smart-meter passage was retrieved but not used by the assessment. Open either example to inspect the exact reasons and passages.

Roles

96VP Product · Quernlade Logisticslogistics softwareTo review
94Principal Product Manager · Madderstow Freightlogistics softwareTo review
92Principal Product Manager · Pellitorby Insightanalytics platformsTo review
91Head of Product, Platforms · Fescuemoor Logisticslogistics softwareTo review
89Product Lead, AI · Dunlinhope Journeystravel platformsTo review
85Senior Director, Product · Smewbeck Studiosmedia and streamingReady to apply
84Group Product Manager · Sallowdene Traveltravel platformsTo review
83Senior Director, Product · Orpinecote Analyticsanalytics platformsReady to apply
83Product Lead, AI · Quernlade Analyticsanalytics platformsReady to apply
82Head of Product, Platforms · Fennelwick Energyenergy softwareReady to apply
82Head of Product, Platforms · Tutsanholt Analyticsanalytics platformsReady to apply
80Product Lead, AI · Corncrakeby Field Servicesfield service softwareTo review
80Senior Director, Product · Quernlade Field Servicesfield service softwareTo review
79Principal Product Manager · Cobbleford Traveltravel platformsTo review
79Principal Product Manager · Tavisholm Freightlogistics softwareReady to apply
78Principal Product Manager · Quernlade Traveltravel platformsTo review
77Principal Product Manager · Charlockholt Analyticsanalytics platformsTo review
76Principal Product Manager · Fescuemoor Mediamedia and streamingTo review
76Principal Product Manager · Quernlade Grid Systemsenergy softwareReady to apply
76Director of Product · Medlarstow Energyenergy softwareReady to apply
74Director, Platform Product · Quernlade Studiosmedia and streamingTo review
71Product Lead, AI · Bistortholt Field Servicesfield service softwareTo review
70Principal Product Manager · Tutsanholt Field Servicesfield service softwareReady to apply
64Head of Product, Platforms · Medlarstow Mediamedia and streamingTo review
61Senior Director, Product · Smewbeck Logisticslogistics softwareTo review
59Head of Product, Platforms · Bistortholt Academylearning technologyTo review
59Product Lead, AI · Smewbeck TransitmobilityTo review
59VP Product · Smewbeck Traveltravel platformsTo review
58Head of Product, Platforms · Corncrakeby Grid Systemsenergy softwareTo review
58Director, Platform Product · Orpinecote Academylearning technologyTo review
57Product Lead, AI · Madderstow Grid Systemsenergy softwareTo review
56Senior Director, Product · Madderstow Learninglearning technologyTo review
52Director of Product · Garganeyholt Learninglearning technologyTo review
51Director, Platform Product · Tavisholm TransitmobilityTo review
50Group Product Manager · Orpinecote Studiosmedia and streamingTo review
44Principal Product Manager · Smewbeck MobilitymobilityNot interested
43Director, Platform Product · Fennelwick TransitmobilityNot interested
43Director of Product · Quernlade TransitmobilityNot interested
41Principal Product Manager · Medlarstow Learninglearning technologyNot interested
39Principal Product Manager · Orpinecote Field Servicesfield service softwareNot interested

96 VP Product · Quernlade Logistics

logistics software · hybrid · found on job feed D 4 weeks ago · status: To review

Why this score

Asks for route and dispatch planning; passage 1 shows it (similarity 0.67).
Asks for team leadership; passage 4 shows it (similarity 0.88).
Asks for analytics and reporting; passage 6 shows it (similarity 0.74).
Asks for shipment tracking; passage 2 shows it (similarity 0.83).

What did not fit

Everything the role asks for has a passage.

Passages the retrieval found

0.88 Led eight product managers and two designers across five teams, with hiring, coaching and a quarterly plan agreed jointly with engineering.
0.83 Owned parcel tracking for a regional carrier: one event history per parcel, scanner updates matched within minutes, and a morning list of stuck parcels for the depot leads.
0.74 Set up product analytics for a fleet app: one naming scheme for events, six agreed measures, and a monthly readout for the leadership team.
0.67 Led the product for a dispatch planner that grouped next-day deliveries into routes each evening; late deliveries fell by a fifth over two quarters.
0.54 Ran retention for a language app: streaks, reminders and a lesson plan that adapted, with the weekly return rate as the measure. (retrieved, not used by the assessment)

An end-to-end candidate experience

Email is a signal. A decision comes next.

Select a stage to inspect the intended handoff and its status against the repository capability inventory. This portfolio preview has no mailbox or calendar connection.

Intended integration flow

Documented laptop intake

Read the signal, preserve the source.

Repository documentation describes a laptop Gmail workflow that imports digest roles and flags interviews, supported by queue and status code. The unified email-to-status MCP flow illustrated here is proposed here; implementation is not established.

Status: laptop intake documented; unified MCP flow not established.

Operational MCP triage is separate from candidate email, calendar and reminders. This preview does not execute tasks or grant integration permissions.

Notes can inform preparation. Source tracking retains notes and evidence. The illustrated feedback loop is intended use; automatic learning or ranking changes remain proposed; implementation is not established.

Read every intended stage and its evidence status

Documented laptop intake

Read the signal, preserve the source.

Status: laptop intake documented; unified MCP flow not established.

Operational MCP triage is separate from candidate email, calendar and reminders. This preview does not execute tasks or grant integration permissions.

Implemented source tracking

Review the recorded application state.

The repository implements status history, notes, evidence and interview flags. Two application lanes retain owner submission: cloud-assisted form filling, or laptop-generated documents for applying on the employer site.

Status: history, notes, evidence and interview flags implemented in source.

The review gate illustrated here is a design proposal. The sample controls above use six local statuses and reset on reload; they do not operate the repository workflow.

Proposed scheduling extension

Turn an invitation into a confirmed plan.

This proposed flow would extract a time and timezone, link the invitation and ask the candidate to confirm before creating an event. The preview shows an illustrative hold only.

Status: proposed here; implementation not established.

The sample calendar has no real date, event or connection. No event is created by this preview.

Proposed follow-through extension

Prompt a deliberate follow-up.

This proposed flow would retain a candidate-chosen follow-up date and notify them to act. Source tracking retains notes and evidence; automated feedback-to-ranking is proposed here; implementation is not established.

Status: proposed here; implementation not established.

No reminder is scheduled in this preview. Recorded notes can inform preparation; they do not establish automatic learning or explain every rejection.

Implementation truth

The product and the direction.

Source status describes repository capabilities, not live deployment or execution. The preserved local model above is an older fictional sample with six status options; the repository inventory has seven visible states.

Capability	Source status	What the evidence supports
Discovery and fit assessment	Source implemented	Multi-source discovery, normalization, deduplication, liveness and LLM fit assessment with strengths and gaps.
Profile and document evidence	Source implemented	Hybrid retrieval finds career evidence for documents. Source-link and verification checks exist with separate shadow/enabled rollout modes. Fit assessment uses an LLM; vector retrieval is not the fit score.
Application lanes and tracking	Source implemented	Two lanes retain owner submission: cloud-assisted form filling, or laptop-generated documents for applying on the employer site. Status history, notes, evidence and interview flags are implemented in source. Complete interview preparation and automatic feedback-to-ranking were not established.
Laptop email / proposed MCP flow	Documented / proposed	Repository documentation describes laptop Gmail intake of digest roles and interview flags, supported by queue/status code. Unified MCP email-to-calendar flow was not established; operational MCP triage is separate.
Calendar and reminders	Proposed	Proposed here; implementation is not established. Calendar holds and reminders above are illustrations only, with no event or reminder created by this preview.

What the history teaches

65 Discovered
39 Reviewed
22 Ready to apply
13 Applied
5 In interview

Three findings from the reports

In the sample history, interviews per tracked role were highest from employer job board A (3 of 5) and lowest from job feed D (0 of 5).
From Reviewed to Ready to apply: 22 of 39 roles; the other 17 were set to Not interested.
From Applied to In interview: 5 of 13 roles.

← Back to About

JobWatch privacy policy