Ashfaque Rifaye

Hiring Agent

Paste a job description. The agent reads the role, maps every requirement to evidence from my work, scores the fit honestly, then drafts your outreach email and a calendar invite. You approve every action.

Built as a product demo: an agent with tools, guardrails and a human in the loop

0 / 8,000

Or try a sample role

The text goes to the site’s AI gateway only to read the requirements and write the brief. Nothing is stored.

The agent’s plan

LLM + rules

  1. Read the roleextract_requirements()LLM

    Extract requirements into a fixed skills vocabulary

  2. Match evidencesearch_evidence()Rules

    Find proof for each requirement in the verified record

  3. Score the fitscore_fit()Rules

    Weighted, explainable score; must-haves count double

  4. Write the briefwrite_brief()LLM

    Summarise using only the matched evidence

  5. Prepare actionsdraft_outreach()Rules

    Draft an email, a calendar invite and a brief

Under the hood: the model perceives, code decides

The same principle behind DeviceFlex. The LLM does what only language models can do; everything that decides an outcome is deterministic and replayable.

  1. 01LLM

    Read the role

    extract_requirements()

    Language is messy; a model maps it onto a fixed vocabulary of 35 skills, and nothing outside that vocabulary survives.

  2. 02Deterministic

    Match evidence

    search_evidence()

    A lookup against 22 verified evidence items. No model, so no invented experience.

  3. 03Deterministic

    Score the fit

    score_fit()

    Weighted and explainable: must-haves count double, and "or" requirements take the stronger match.

  4. 04LLM

    Write the brief

    write_brief()

    The model sees only the matched evidence and the computed score, which it may not change.

  5. 05Deterministic

    Prepare actions

    draft_outreach()

    Allowlisted outputs only. The email and invite are drafts you send yourself.

  • Evidence or nothingEvery claim links to a case study, role or build. Gaps are reported plainly.
  • Injection-resistantThe job description is treated as data. Actions come from a fixed allowlist, never from the text.
  • Human in the loopNothing is sent, booked or saved until you click.
  • Graceful without the modelIf the gateway is down, a rules-based reader and template writer take over.

Building an AI product, scaling an enterprise AI experience, or figuring out where AI actually creates value?

I work at the intersection of product strategy, AI systems and execution.

Let’s talkOther ways to reach me

Direct