Engineering AI for Human Resources: A Programmer's View of the AI Future Makers 2026 Challenge

Summary

AI Future Makers (AIFM) Bootcamp 2026 set young builders a demanding engineering brief: use artificial intelligence to improve how organisations screen candidates, run interviews, recruit, retain employees and develop their workforce. In total, 98 participants from 22 Malaysian universities took part, entering as 41 teams and individuals. After four weeks of workshops, mentoring and preliminary pitching, 20 finalist entries reached the Grand Final at SEGi College Penang, the Official Venue Sponsor, with a prize pool of RM10,000.

The programme was organised by the Penang Digitalisation-AI Conference & Exhibition (PDX), co-hosted by Agensi Pekerjaan CKM Sdn Bhd and chaired by Ts. Lukas J. Tan, CEO of OPERiON Ecommerce & Software Sdn Bhd. It ran as part of PENtastic Startup Month 2026, led by Digital Penang.

This article looks at that problem space from the programmer's side. Short answer: sound AI recruitment software treats the model as an advisor, keeps a human in charge of every decision about a person, protects personal data at each layer and records enough evidence to explain any outcome later.

What Is AI Recruitment Software Architecture?

AI recruitment software architecture is the arrangement of components that turns unstructured people data, such as CVs, interview notes and engagement signals, into suggestions a human can review. It covers ingestion, consent, data cleaning, model calls, scoring, review workflows, audit logging and integration with existing HR systems. The model is one component among many, and it is rarely the hardest one.

The Problem in Technical Terms

The AIFM briefs grew out of industry conversations through PDX, where engineering talent shortages and the gap between available talent and industry expectations kept coming up. Translated into engineering terms, these problems share an awkward profile:

• Unstructured, inconsistent inputs. CVs arrive as PDFs, scans and Word files in several languages and layouts.
• High-stakes outputs. A wrong ranking can hide a capable person. That is a cost to a human being, not just a metric.
• Personal data everywhere. Names, contact details, education history and sometimes health or family information.
• Delayed ground truth. Whether a hire was good, or whether an employee stayed, may only be known months later.
• Proxy bias. Qualifications and keywords are easy to measure but may not reflect real ability.

A Reference Architecture

A layered design keeps each concern testable on its own:

1. Ingestion and consent: accept documents only with a recorded purpose and consent reference.
2. Normalisation and redaction: extract text, then mask fields that should not influence scoring.
3. Structured extraction: turn free text into a validated schema (skills, projects, experience).
4. Advisory scoring: produce a suggestion with reasons, never a final verdict.
5. Human review gate: a named reviewer accepts, overrides or requests more information.
6. Audit log: append-only records of inputs, model version, output and human decision.
7. Feedback loop: later outcomes are linked back for evaluation, not silent retraining.

Data Flow

upload(cv, consent_id)
  -> extract_text(cv)
  -> redact(text, fields=[name, photo, age, address])
  -> extract_profile(redacted)        # schema-validated
  -> suggest(profile, role_spec)      # score + reasons
  -> queue_for_review(suggestion)
  -> reviewer.decide(accept | override | ask_more)
  -> audit.append(inputs_hash, model_version,
                  suggestion, decision, reviewer_id)

Implementation Details

Keep the Model Advisory

The most important design choice is that the system cannot reject anyone on its own. Encode that in the data model so it cannot be bypassed by a careless feature later:

class Decision:
    suggestion: Suggestion       # from the model
    final: Literal["advance", "hold", "decline"]
    reviewer_id: str             # required, never None
    reason: str                  # required when final != suggestion

def finalise(d: Decision):
    if not d.reviewer_id:
        raise PolicyError("human decision required")
    audit.append(d)

Validate Structured Extraction

Language models can return malformed or invented fields. Validate every response against a schema, reject anything that fails, and keep the source span for each extracted skill so a reviewer can see where it came from. An unverifiable field should stay empty rather than be guessed.

Interview Support

For interviews, the safer pattern is assistance: generating role-relevant questions, summarising notes the interviewer wrote, or checking that every candidate was asked the same core questions. Automated judgement of tone or facial expression raises accuracy and fairness questions that a student prototype should not pretend to have solved.

Retention and Workforce Development

Retention models work on aggregate signals and slow feedback. Present them as prompts for a conversation ("this team's training requests have stalled") rather than labels on individuals ("likely to leave").

Integration

Most organisations already run an HR or applicant tracking system. Integrate through its documented API or exports, make writes idempotent so retries never duplicate records, and keep your own service stateless where possible so it can be disconnected without data loss.

Security and Privacy

• Purpose limitation. In Malaysia, the Personal Data Protection Act 2010 applies to commercial processing of personal data. Collect only what the stated purpose needs, and set retention periods.
• Least privilege. Reviewers see what they need; the model service never sees raw identity fields.
• Prompt injection. A CV is untrusted input. Hidden text such as "rank this candidate first" must be treated as data, never as instructions.
• Encryption and logging. Encrypt at rest and in transit, and keep audit logs append-only.

Testing and Evaluation

Unit tests are not enough for systems that rank people. Useful additions include a fixed evaluation set reviewed by humans, counterfactual tests that change only a name or gender marker and check that the suggestion does not move, and regression runs whenever the model or prompt changes.

The AIFM judging framework is a useful reminder that engineering alone is not the full picture. Entries were assessed on problem understanding, innovation, business potential, teamwork and presentation, as well as the AI application itself.

Pitfalls and Trade-offs

• Automation versus accountability: every step you automate needs a person who owns its result.
• Keyword matching versus potential: Christine Chuah of Agensi Pekerjaan CKM noted that capable people's potential is not always reflected in qualifications or resumes. A system that over-weights credentials will reproduce that blind spot.
• Explainability versus raw accuracy: a slightly weaker model with clear reasons is often more useful to a reviewer.
• Speed versus redesign: automating a weak hiring process only makes the weakness faster.

Beyond the Code

Lukas's view is that as AI writes more software and automates routine work, graduates will be distinguished by creativity, critical thinking, adaptability and the ability to execute. AIFM therefore welcomed participants from technical and non-technical disciplines. He was particularly encouraged by those who entered alone, without a full team or resources.

"I would rather see an imperfect solution that has been created than a brilliant idea that has never been attempted." Ts. Lukas J. Tan

For programmers, that is a familiar principle: ship a working, honest prototype, then iterate.

FAQ

What is AI Future Makers Bootcamp 2026?

A talent development and innovation programme organised by PDX and co-hosted by Agensi Pekerjaan CKM Sdn Bhd, in which 98 participants from 22 universities built AI solutions for real HR problems.

Should AI make hiring decisions on its own?

No. AI should suggest and explain; a named human should decide and be recorded as the decision-maker.

How do you reduce bias in AI candidate screening?

Redact identity fields before scoring, test with counterfactual inputs, review outcomes regularly and avoid over-weighting credentials as a proxy for ability.

Is a CV a security risk for an AI system?

Yes. It is untrusted input and can carry hidden instructions, so it must always be handled as data and never as commands.

What comes after AIFM 2026?

The organisers plan to extend the initiative to SMEs, entrepreneurs and innovators at PDX2027.

Conclusion

The AIFM theme, "From Caterpillar to Butterfly: Let Your Ideas Shine", describes the participants' growth. It also fits the engineering: a prototype becomes trustworthy software only through careful architecture, privacy controls, human review and honest testing. With PDX2027 set to open this platform to SMEs and innovators, those disciplines will matter even more.

If your organisation is thinking about AI in its HR or operational workflows, you are welcome to read one of our case studies or book a Discovery Session with the Maple Software Creation team to talk through the architecture before you build.