Haris MekicIndependent candidate portfolioA technical portfolio about turning real work into AI-assisted workflows—and keeping evidence, permissions and review in the loop.
FROM MY OWN WORKING PRACTICE
I built CareerAgent to keep source evidence and career context connected across repeated CV and application drafts—so a new task starts from the underlying records, not a short CV that invites unsupported claims.
Career documents, operational records and verified vacancies become reusable claims, tailored documents and application case state. I direct AI-assisted implementation with Claude and Codex and check the results against sources. Python, SQLite and source hashes make the state inspectable.
IMPLEMENTED PERSONAL WORKFLOW · VERIFIED · EPOCH-AI-004
This Epoch application is being prepared through that working method: source inspection, a tailored CV, technical answers and claim checks. Submission remains a separate controlled step. EPOCH-AI-008 · VERIFIED PREPARATION
This interactive replay explains the implemented workflow using an illustrative record. It runs locally in your browser; no AI is called and no application is submitted.
Lexical retrieval—not vector retrieval. Separate workflow commits—not one distributed ACID transaction. Evidence of implementation and personal use—not a measured hiring outcome, time saving or commercial rollout.
A Python control layer, model-provider interface and file-backed working state, with explicit rules for proceeding, reporting back or queuing review. High-risk work queues for human review.
LOCAL IMPLEMENTATION · VERIFIED · EPOCH-AI-003
Specialist analysis shares one context, separates business fit, staffing and compliance, then feeds a dependent operations recommendation with complete, partial or failed status.
SOURCE-VERIFIED IMPLEMENTATION · EPOCH-AI-002
Explored local pretrained-model adaptation with LoRA. The saved run is an experiment in a bounded method—not a claim of foundation-model training or proven task-quality gains.
SAVED EXPERIMENT · VERIFIED ARTIFACT · EPOCH-AI-006
Individual coaching on the working method. One participant independently tried a hypothetical consulting assignment and reported practical value; I followed with a proposal for a supervised real-work pilot.
INDIVIDUAL HYPOTHETICAL TRIAL · HIGH CONFIDENCE · EPOCH-AI-005
A PROPOSED APPROACH, NOT A DIAGNOSIS
Listen first. Find one recurring task. Establish what good looks like. Make the smallest safe pilot.
I would start with the team’s actual priorities, existing tools and constraints. These are prospective contributions—not assertions that Epoch has any particular problem.
Understand research-support work, handoffs and friction with the people doing it. Ask what must stay human and what would make the work easier.
Observe one repeatable task before changing it. Record quality criteria, review burden and an honest baseline; avoid invented savings.
Use approved non-sensitive material for a research-support task—such as source triage or a structured evidence summary—with permissions and a named reviewer.
Document context, failure cases, ownership and rollback. Coach teammates through the method, and agree who checks and maintains the workflow.
Edit the assumptions. The result is a transparent discussion aid, not an AI recommendation or authorization to run anything.
| Candidate task | Value | Confidence | Effort (h) | Risk |
|---|
Priority = (value × confidence) ÷ effort. Value and confidence are bounded to 1–5; effort to 1–40 hours; budget to 0–80 hours. Ties prefer lower effort, then the original task order.
Without approved access, data handling and a named reviewer, all work is held. High-risk tasks always require separate specialist approval and cannot become pilot-ready through this calculator. Other tasks are selected in score order only if they fit the remaining budget. Every output still needs human review. No task is activated.
IF THE FIRST PILOT EARNS ITS PLACE
Understand the workflow, agree a baseline and test a narrow research-support pilot with documented review.
Compare against the baseline, examine failure cases, iterate on context and help colleagues use the method.
Keep only what proves useful. Assign ownership, monitor quality and review burden, and retire weak experiments.
INSPECTABLE, WITH LIMITATIONS
This page separates what exists, what has been observed and what I would propose.
Public records are deliberately sanitized. Source paths, private application records, client information and participant identities are not published. Confidence labels describe the reviewed claim—not a guarantee of future performance.
Five distinct system-family entries were assessed in a September source review; three were freshly source-spotchecked. These are not five unrelated or deployed products. Four selected applied cases include individual coaching; it is not an additional software system.
Two SQLite initialization schema definitions were inspected. This is source evidence—not a count of live databases, rows or the final migrated schema.
A provider-usage audit covers four selected local sessions: three Codex and one Claude. Codex maximum cumulative counters are counted once per session, and cached input is a subset of input. Claude message IDs are deduplicated; cache reads and creation remain separate. Replayed context, tools and inherited state can affect counters. These are not human-typed tokens or all-time usage.
Inspect the bounded usage audit ↗ · Inspect assessment coverage ↗
No verified human-hours, total cost, aggregate tokens, adoption or ROI figure is claimed here. Reviewed on 14 October 2026.